{ "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": 0.025568461045622826, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.025274718180298805, "qparams": { "group_size": 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": 0.0118113923817873, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.011752965860068798, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.011687000282108784, "qparams": { "group_size": 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": 0.007828956469893456, "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": 0.0402623750269413, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.025021834298968315, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.011710016056895256, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.011655948124825954, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.012818994000554085, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.014789460226893425, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01163862831890583, "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": 0.008687952533364296, "qparams": { "group_size": 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": 0.007801939267665148, "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": 0.009514601901173592, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.007794337347149849, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.007539526093751192, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.007792765274643898, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.007537690456956625, "qparams": { "group_size": 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": 0.008521204814314842, "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": 0.007792074233293533, "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": 0.008201991207897663, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.007536846213042736, "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": 0.025568461045622826, "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": 0.025568461045622826, "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": 0.025063930079340935, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.024712132290005684, "qparams": { "group_size": 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": 0.011954691261053085, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.01187711488455534, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.011799891479313374, "qparams": { "group_size": 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": 0.006220320239663124, "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": 0.032342128455638885, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.024449242278933525, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.011838696897029877, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.011748230084776878, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.01194845512509346, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.01281377486884594, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.011726345866918564, "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": 0.007542649284005165, "qparams": { "group_size": 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": 0.006174369249492884, "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": 0.007343755569308996, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0061551895923912525, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.005629455670714378, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.006151419598609209, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.005625098943710327, "qparams": { "group_size": 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": 0.005623807664960623, "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": 0.006150617729872465, "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": 0.004933486692607403, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005622015334665775, "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": 0.025063930079340935, "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": 0.025063930079340935, "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": 0.07145129144191742, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.05957668647170067, "qparams": { "group_size": 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": 0.04534521698951721, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.038464076817035675, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.031761396676301956, "qparams": { "group_size": 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": 0.021835261955857277, "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": 0.0624033622443676, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04752092435956001, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03385717049241066, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.02625022456049919, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02732979692518711, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03309497982263565, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022982360795140266, "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": 0.015836572274565697, "qparams": { "group_size": 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": 0.013657418079674244, "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": 0.01645316742360592, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009015306830406189, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.007984308525919914, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.008063534274697304, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.006738285068422556, "qparams": { "group_size": 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": 0.00846775434911251, "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": 0.007537518162280321, "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": 0.0047459895722568035, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0054654935374855995, "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": 0.07145129144191742, "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": 0.07145129144191742, "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": 0.025073684751987457, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.015645086765289307, "qparams": { "group_size": 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": 0.010055646300315857, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.010741518810391426, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.010263794101774693, "qparams": { "group_size": 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": 0.004990055225789547, "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": 0.015839675441384315, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.014040576294064522, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.011632366105914116, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.007250404916703701, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.007600197568535805, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.008039598353207111, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.00700798723846674, "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": 0.005428828299045563, "qparams": { "group_size": 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": 0.005000438075512648, "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": 0.004257958382368088, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.003631957806646824, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0035035244654864073, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.003258131444454193, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0030432231724262238, "qparams": { "group_size": 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": 0.002702871337532997, "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": 0.003233409719541669, "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": 0.002296553459018469, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0028745997697114944, "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": 0.025073684751987457, "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": 0.025073684751987457, "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": 0.08593808114528656, "qparams": { "group_size": 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": 0.07946498692035675, "qparams": { "group_size": 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": 0.07736273854970932, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.06916076689958572, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.038582902401685715, "qparams": { "group_size": 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": 0.03651360049843788, "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": 0.04329993575811386, "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": 0.04007468745112419, "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": 0.03920309245586395, "qparams": { "group_size": 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": 0.034465909004211426, "qparams": { "group_size": 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": 0.032441917806863785, "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": 0.02201901376247406, "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": 0.019325191155076027, "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": 0.018684357404708862, "qparams": { "group_size": 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": 0.018537990748882294, "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": 0.011104225181043148, "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": 0.010258237831294537, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010208414867520332, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.009533103555440903, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009438655339181423, "qparams": { "group_size": 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": 0.006267664022743702, "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": 0.007092497777193785, "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": 0.0060551282949745655, "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": 0.005669076927006245, "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": 0.08593808114528656, "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": 0.08593808114528656, "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": 0.13025911152362823, "qparams": { "group_size": 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": 0.12056026607751846, "qparams": { "group_size": 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": 0.11749029904603958, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10503941029310226, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.05839184671640396, "qparams": { "group_size": 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": 0.05535397678613663, "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": 0.06548206508159637, "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": 0.06045518442988396, "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": 0.05934271216392517, "qparams": { "group_size": 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": 0.05214984342455864, "qparams": { "group_size": 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": 0.04906085506081581, "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": 0.03301766514778137, "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": 0.028697140514850616, "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": 0.02780955471098423, "qparams": { "group_size": 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": 0.027592575177550316, "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": 0.016480019316077232, "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": 0.01447820384055376, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014407600276172161, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013277475722134113, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013141784816980362, "qparams": { "group_size": 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": 0.008777947165071964, "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": 0.00904623232781887, "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": 0.008464506827294827, "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": 0.006319723092019558, "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": 0.12056026607751846, "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": 0.12056026607751846, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.07253167033195496, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.05899422615766525, "qparams": { "group_size": 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": 0.04031582549214363, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.036384761333465576, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.030058477073907852, "qparams": { "group_size": 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": 0.020053822547197342, "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": 0.05745183303952217, "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": 0.04907640069723129, "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": 0.032522544264793396, "qparams": { "group_size": 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": 0.025524262338876724, "qparams": { "group_size": 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": 0.025679033249616623, "qparams": { "group_size": 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": 0.025837203487753868, "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": 0.022916624322533607, "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": 0.016000228002667427, "qparams": { "group_size": 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": 0.014475683681666851, "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": 0.014112895354628563, "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": 0.011325865052640438, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.00982329249382019, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.010758337564766407, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.008964572101831436, "qparams": { "group_size": 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": 0.010337159037590027, "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": 0.010463553480803967, "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": 0.008391303941607475, "qparams": { "group_size": 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": 0.008272376842796803, "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": 0.07253167033195496, "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": 0.07253167033195496, "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": 0.018364280462265015, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.014532212167978287, "qparams": { "group_size": 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": 0.011054065078496933, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.009697607718408108, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.007857311517000198, "qparams": { "group_size": 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": 0.00520723219960928, "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": 0.014276677742600441, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.011436882428824902, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.008567607961595058, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.006449665874242783, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.006856578402221203, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.007571343332529068, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0056454394944012165, "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": 0.003913013264536858, "qparams": { "group_size": 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": 0.0033442028798162937, "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": 0.0038899793289601803, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.002185315592214465, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0019279144471511245, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.001942712813615799, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0015935349510982633, "qparams": { "group_size": 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": 0.0020469347946345806, "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": 0.0018134956480935216, "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": 0.0012192234862595797, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0012679194333031774, "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": 0.018364280462265015, "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": 0.018364280462265015, "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": 0.017255345359444618, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.013384321704506874, "qparams": { "group_size": 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": 0.01022270042449236, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.008943186141550541, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.007193946745246649, "qparams": { "group_size": 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": 0.004874942358583212, "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": 0.011839167214930058, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.01042142789810896, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.007938412018120289, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.005779809318482876, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.00588923878967762, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0059866951778531075, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.00498950993642211, "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": 0.003553033107891679, "qparams": { "group_size": 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": 0.003125346265733242, "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": 0.0030084962490946054, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0019980533979833126, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.001819778117351234, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0017534141661599278, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0014910635072737932, "qparams": { "group_size": 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": 0.001612199004739523, "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": 0.0016281442949548364, "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": 0.0011027391301468015, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0011819442734122276, "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": 0.017255345359444618, "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": 0.017255345359444618, "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": 0.09155873209238052, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0765240490436554, "qparams": { "group_size": 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": 0.06999721378087997, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.05975433066487312, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04079989343881607, "qparams": { "group_size": 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": 0.034492217004299164, "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": 0.051139235496520996, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04691826179623604, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.042993176728487015, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.032660964876413345, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.030482428148388863, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0261144507676363, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02249758131802082, "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": 0.019724305719137192, "qparams": { "group_size": 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": 0.019020162522792816, "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": 0.01310260035097599, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010426667518913746, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010162921622395515, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.00885712169110775, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.008383666165173054, "qparams": { "group_size": 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": 0.006929048337042332, "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": 0.007060510106384754, "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": 0.005960783455520868, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0049063279293477535, "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": 0.09155873209238052, "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": 0.09155873209238052, "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": 0.1678941547870636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10428594052791595, "qparams": { "group_size": 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": 0.06913862377405167, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07228157669305801, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07075446844100952, "qparams": { "group_size": 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": 0.040146131068468094, "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": 0.10352370142936707, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09245409071445465, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08055698871612549, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.047058749943971634, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.049919482320547104, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05351291969418526, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04564741998910904, "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": 0.03610434755682945, "qparams": { "group_size": 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": 0.03347473219037056, "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": 0.027457784861326218, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.021702153608202934, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020944803953170776, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01807103306055069, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01648123562335968, "qparams": { "group_size": 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": 0.015913590788841248, "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": 0.01788630150258541, "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": 0.013011648319661617, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014820721000432968, "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": 0.10428594052791595, "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": 0.10428594052791595, "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": 0.1300843209028244, "qparams": { "group_size": 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": 0.1221398115158081, "qparams": { "group_size": 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": 0.11983245611190796, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10807807743549347, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.057991527020931244, "qparams": { "group_size": 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": 0.055715594440698624, "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": 0.0641821026802063, "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": 0.0593063123524189, "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": 0.058655738830566406, "qparams": { "group_size": 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": 0.052741050720214844, "qparams": { "group_size": 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": 0.04996054247021675, "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": 0.03251294046640396, "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": 0.028327396139502525, "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": 0.02777138352394104, "qparams": { "group_size": 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": 0.027636853978037834, "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": 0.016316108405590057, "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": 0.01485004648566246, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014813653193414211, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01389880571514368, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013820057734847069, "qparams": { "group_size": 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": 0.00903545506298542, "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": 0.00976473931223154, "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": 0.008859154768288136, "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": 0.007455968763679266, "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": 0.1221398115158081, "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": 0.1221398115158081, "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": 0.16819432377815247, "qparams": { "group_size": 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": 0.15849527716636658, "qparams": { "group_size": 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": 0.15560859441757202, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1408797651529312, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.07521344721317291, "qparams": { "group_size": 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": 0.0723767876625061, "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": 0.08298047631978989, "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": 0.07678314298391342, "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": 0.07604037970304489, "qparams": { "group_size": 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": 0.0686006098985672, "qparams": { "group_size": 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": 0.0649566575884819, "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": 0.04168492928147316, "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": 0.03625142574310303, "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": 0.03556511178612709, "qparams": { "group_size": 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": 0.03540295362472534, "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": 0.02074000984430313, "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": 0.018079262226819992, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0180378258228302, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.016772300004959106, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01667102985084057, "qparams": { "group_size": 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": 0.010754215531051159, "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": 0.010591845959424973, "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": 0.010522596538066864, "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": 0.0066873421892523766, "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": 0.08298047631978989, "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": 0.08298047631978989, "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": 0.14703604578971863, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.13187482953071594, "qparams": { "group_size": 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": 0.1199994683265686, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10837417840957642, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.06630116701126099, "qparams": { "group_size": 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": 0.056368619203567505, "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": 0.08655692636966705, "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": 0.07961136102676392, "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": 0.06839524209499359, "qparams": { "group_size": 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": 0.05900296941399574, "qparams": { "group_size": 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": 0.05718885734677315, "qparams": { "group_size": 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": 0.04350728169083595, "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": 0.040249988436698914, "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": 0.0322912372648716, "qparams": { "group_size": 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": 0.030265942215919495, "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": 0.022420555353164673, "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": 0.018104033544659615, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.016885695978999138, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01684102602303028, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015333233401179314, "qparams": { "group_size": 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": 0.013158062472939491, "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": 0.013836788944900036, "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": 0.011004538275301456, "qparams": { "group_size": 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": 0.010114429518580437, "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": 0.1199994683265686, "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": 0.1199994683265686, "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": 0.03880651667714119, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.03525949642062187, "qparams": { "group_size": 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": 0.03364023566246033, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.029681570827960968, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.017550650984048843, "qparams": { "group_size": 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": 0.016086673364043236, "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": 0.022790128365159035, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.01909574866294861, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.017941968515515327, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0153003865852952, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.014471870847046375, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.011573443189263344, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.009168798103928566, "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": 0.008419030345976353, "qparams": { "group_size": 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": 0.008228101767599583, "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": 0.005805565509945154, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.004397585988044739, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.004312272649258375, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.003959068097174168, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.003838919335976243, "qparams": { "group_size": 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": 0.00304411631077528, "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": 0.0028455485589802265, "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": 0.0025730235502123833, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0019494867883622646, "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": 0.03880651667714119, "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": 0.03880651667714119, "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": 0.03146396204829216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.02848135493695736, "qparams": { "group_size": 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": 0.027183497324585915, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.023926734924316406, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.014157389290630817, "qparams": { "group_size": 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": 0.012996120378375053, "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": 0.01683032140135765, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.015376039780676365, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.01449219323694706, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.012291720137000084, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.011517408303916454, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.008480013348162174, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.007321961689740419, "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": 0.006778491660952568, "qparams": { "group_size": 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": 0.0066489954479038715, "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": 0.0042479210533201694, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0035344595089554787, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.00347887328825891, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0031698187813162804, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0030858702957630157, "qparams": { "group_size": 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": 0.002246295800432563, "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": 0.002263447502627969, "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": 0.0020521990954875946, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0015469634672626853, "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": 0.03146396204829216, "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": 0.03146396204829216, "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": 0.14847683906555176, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13694171607494354, "qparams": { "group_size": 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": 0.13320589065551758, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11879809200763702, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0681936964392662, "qparams": { "group_size": 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": 0.06422185897827148, "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": 0.07670776546001434, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0709148496389389, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06944210827350616, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06025785207748413, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05647572875022888, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03885747492313385, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0337531603872776, "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": 0.03252844512462616, "qparams": { "group_size": 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": 0.032239820808172226, "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": 0.019378390163183212, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.016575299203395844, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.016485443338751793, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.014960742555558681, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014774118550121784, "qparams": { "group_size": 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": 0.010003028437495232, "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": 0.009840907528996468, "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": 0.009564983658492565, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006177624221891165, "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": 0.11879809200763702, "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": 0.11879809200763702, "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": 0.18453185260295868, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14938755333423615, "qparams": { "group_size": 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": 0.1352408230304718, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10882477462291718, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0824379026889801, "qparams": { "group_size": 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": 0.06864053755998611, "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": 0.10370488464832306, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09371314942836761, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08754255622625351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06111845374107361, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0569281242787838, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05355307087302208, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04594498127698898, "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": 0.04091983661055565, "qparams": { "group_size": 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": 0.039657097309827805, "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": 0.02749469503760338, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02332771010696888, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022964639589190483, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01972644031047821, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01894836314022541, "qparams": { "group_size": 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": 0.0161548163741827, "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": 0.017491135746240616, "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": 0.014700915664434433, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014303000643849373, "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": 0.10882477462291718, "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": 0.10882477462291718, "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": 0.18454349040985107, "qparams": { "group_size": 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": 0.174002543091774, "qparams": { "group_size": 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": 0.17084617912769318, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1552543193101883, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08475141227245331, "qparams": { "group_size": 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": 0.08142418414354324, "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": 0.09360656887292862, "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": 0.08648983389139175, "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": 0.08568719774484634, "qparams": { "group_size": 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": 0.0771305188536644, "qparams": { "group_size": 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": 0.0734727755188942, "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": 0.0475703626871109, "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": 0.04128213971853256, "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": 0.040500346571207047, "qparams": { "group_size": 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": 0.040322933346033096, "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": 0.023810580372810364, "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": 0.021099723875522614, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021056674420833588, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019639965146780014, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01952659711241722, "qparams": { "group_size": 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": 0.012821679934859276, "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": 0.013111754320561886, "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": 0.012580462731420994, "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": 0.009252442978322506, "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": 0.09360656887292862, "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": 0.09360656887292862, "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": 0.2230731099843979, "qparams": { "group_size": 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": 0.21050214767456055, "qparams": { "group_size": 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": 0.20686346292495728, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18819758296012878, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10258770734071732, "qparams": { "group_size": 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": 0.09866874665021896, "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": 0.11326403170824051, "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": 0.10469068586826324, "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": 0.10374893248081207, "qparams": { "group_size": 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": 0.0935029610991478, "qparams": { "group_size": 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": 0.08899692445993423, "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": 0.0572914257645607, "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": 0.04975401237607002, "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": 0.048829302191734314, "qparams": { "group_size": 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": 0.04861130192875862, "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": 0.028606992214918137, "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": 0.02487931214272976, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.024818172678351402, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023044727742671967, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022910226136446, "qparams": { "group_size": 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": 0.014969185926020145, "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": 0.014611595310270786, "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": 0.01467221975326538, "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": 0.00928211584687233, "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": 0.11326403170824051, "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": 0.11326403170824051, "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": 0.21633291244506836, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.19525200128555298, "qparams": { "group_size": 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": 0.18739496171474457, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.16918456554412842, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09728831797838211, "qparams": { "group_size": 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": 0.08974049240350723, "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": 0.11334257572889328, "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": 0.10424263775348663, "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": 0.10027744621038437, "qparams": { "group_size": 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": 0.08592666685581207, "qparams": { "group_size": 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": 0.08187174797058105, "qparams": { "group_size": 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": 0.05744696035981178, "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": 0.04974540323019028, "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": 0.04664260521531105, "qparams": { "group_size": 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": 0.04589622840285301, "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": 0.028835901990532875, "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": 0.024459488689899445, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.024299845099449158, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02226783148944378, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021803488954901695, "qparams": { "group_size": 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": 0.015587392263114452, "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": 0.015757348388433456, "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": 0.014604227617383003, "qparams": { "group_size": 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": 0.011094690300524235, "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": 0.11334257572889328, "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": 0.11334257572889328, "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": 0.053630948066711426, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.04877712205052376, "qparams": { "group_size": 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": 0.046698663383722305, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.04150759056210518, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.02434387616813183, "qparams": { "group_size": 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": 0.02246900647878647, "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": 0.029636356979608536, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.02636932209134102, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.025168828666210175, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.021416179835796356, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02032945677638054, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.015210305340588093, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.012665711343288422, "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": 0.011687835678458214, "qparams": { "group_size": 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": 0.011437389068305492, "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": 0.007673285901546478, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.006088976748287678, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.006013651378452778, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.005510836374014616, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.005354105960577726, "qparams": { "group_size": 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": 0.004056383855640888, "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": 0.0039055999368429184, "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": 0.003582240082323551, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0026497854851186275, "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": 0.053630948066711426, "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": 0.053630948066711426, "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": 0.04338616132736206, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.039704740047454834, "qparams": { "group_size": 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": 0.03793521225452423, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.03369726240634918, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.01977108046412468, "qparams": { "group_size": 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": 0.018275946378707886, "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": 0.023569904267787933, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.021481990814208984, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.020212272182106972, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.017362913116812706, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.016454458236694336, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.011922918260097504, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.010245736688375473, "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": 0.009504628367722034, "qparams": { "group_size": 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": 0.009325122460722923, "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": 0.00597765576094389, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.004987939726561308, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.004913500510156155, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.004521356429904699, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.004407177213579416, "qparams": { "group_size": 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": 0.0031777063850313425, "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": 0.0032371392007917166, "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": 0.002921451348811388, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0022717725951224566, "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": 0.04338616132736206, "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": 0.04338616132736206, "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": 0.19144108891487122, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17738549411296844, "qparams": { "group_size": 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": 0.1728568971157074, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1548222154378891, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.08873523026704788, "qparams": { "group_size": 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": 0.0838170275092125, "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": 0.1002826988697052, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09199691563844681, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09027022123336792, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07869953662157059, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07416339218616486, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.050861433148384094, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04386569559574127, "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": 0.042390480637550354, "qparams": { "group_size": 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": 0.042033400386571884, "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": 0.025405915454030037, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.021603772416710854, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02149197645485401, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019552219659090042, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.019327973946928978, "qparams": { "group_size": 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": 0.013167839497327805, "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": 0.01280822604894638, "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": 0.012602005153894424, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008045869879424572, "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": 0.1002826988697052, "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": 0.1002826988697052, "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": 0.14614595472812653, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1250748485326767, "qparams": { "group_size": 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": 0.1140139102935791, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09118495136499405, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06671805679798126, "qparams": { "group_size": 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": 0.05709211900830269, "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": 0.09106060862541199, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07771670818328857, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06909684091806412, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.052969854325056076, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.048198312520980835, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.046639762818813324, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.038939788937568665, "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": 0.03407672047615051, "qparams": { "group_size": 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": 0.03278682380914688, "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": 0.024855012074112892, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02070946991443634, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02020467258989811, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.018550870940089226, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.017834730446338654, "qparams": { "group_size": 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": 0.015460079535841942, "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": 0.016679098829627037, "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": 0.013802901841700077, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014380418695509434, "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": 0.1140139102935791, "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": 0.1140139102935791, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.200906440615654, "qparams": { "group_size": 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": 0.18898305296897888, "qparams": { "group_size": 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": 0.1853947788476944, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.16828389465808868, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09401185065507889, "qparams": { "group_size": 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": 0.09008395671844482, "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": 0.10403304547071457, "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": 0.09623894095420837, "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": 0.09521140903234482, "qparams": { "group_size": 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": 0.0851413682103157, "qparams": { "group_size": 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": 0.08085175603628159, "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": 0.05331706628203392, "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": 0.046482574194669724, "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": 0.04551141336560249, "qparams": { "group_size": 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": 0.045282505452632904, "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": 0.026758940890431404, "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": 0.02438487857580185, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.024327823892235756, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02271035872399807, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022575795650482178, "qparams": { "group_size": 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": 0.01475259568542242, "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": 0.016066795215010643, "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": 0.014449618756771088, "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": 0.012290526181459427, "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": 0.10403304547071457, "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": 0.10403304547071457, "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": 0.24624837934970856, "qparams": { "group_size": 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": 0.23180744051933289, "qparams": { "group_size": 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": 0.2274840772151947, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20653745532035828, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11499889940023422, "qparams": { "group_size": 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": 0.11019936203956604, "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": 0.12724116444587708, "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": 0.1176493689417839, "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": 0.11645052582025528, "qparams": { "group_size": 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": 0.1041199192404747, "qparams": { "group_size": 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": 0.09889355301856995, "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": 0.06479969620704651, "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": 0.05619809776544571, "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": 0.0550420805811882, "qparams": { "group_size": 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": 0.05477583780884743, "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": 0.03235931321978569, "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": 0.028424028307199478, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02834487147629261, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026246899738907814, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026067720726132393, "qparams": { "group_size": 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": 0.01717260107398033, "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": 0.017262380570173264, "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": 0.016786765307188034, "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": 0.011659115552902222, "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": 0.1176493689417839, "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": 0.1176493689417839, "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": 0.043313540518283844, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.039363663643598557, "qparams": { "group_size": 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": 0.027247756719589233, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.024537401273846626, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.018519865348935127, "qparams": { "group_size": 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": 0.0120896240696311, "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": 0.01933806575834751, "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": 0.032146163284778595, "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": 0.019046474248170853, "qparams": { "group_size": 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": 0.016803987324237823, "qparams": { "group_size": 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": 0.01632835902273655, "qparams": { "group_size": 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": 0.015257781371474266, "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": 0.01423796359449625, "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": 0.006591325160115957, "qparams": { "group_size": 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": 0.006786877289414406, "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": 0.004338255152106285, "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": 0.004137850366532803, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0035655794199556112, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0037727702874690294, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.003091593971475959, "qparams": { "group_size": 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": 0.0031957244500517845, "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": 0.0032181469723582268, "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": 0.0023078990634530783, "qparams": { "group_size": 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": 0.0019350486109033227, "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": 0.043313540518283844, "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": 0.043313540518283844, "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": 0.06951987743377686, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.06331324577331543, "qparams": { "group_size": 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": 0.0607142373919487, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.05385454371571541, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03177125006914139, "qparams": { "group_size": 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": 0.02935209311544895, "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": 0.03702013939619064, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03413151949644089, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03249005973339081, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.027739234268665314, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02604602836072445, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.018743246793746948, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.016268393024802208, "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": 0.01522462535649538, "qparams": { "group_size": 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": 0.014970144256949425, "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": 0.009369234554469585, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.007887929677963257, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.007788689807057381, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.007091360632330179, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.006928093731403351, "qparams": { "group_size": 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": 0.0048992750234901905, "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": 0.004963558167219162, "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": 0.004545391071587801, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003327959682792425, "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": 0.06951987743377686, "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": 0.06951987743377686, "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": 0.06141755357384682, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.055891454219818115, "qparams": { "group_size": 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": 0.053490061312913895, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.047440119087696075, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.028044678270816803, "qparams": { "group_size": 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": 0.025846658274531364, "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": 0.032976772636175156, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.030291715636849403, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.028684956952929497, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.02446122281253338, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02305741049349308, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.01669219881296158, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01442486047744751, "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": 0.013441241346299648, "qparams": { "group_size": 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": 0.013198882341384888, "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": 0.008345469832420349, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0069518969394266605, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.006853712722659111, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.006241865921765566, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0060865916311740875, "qparams": { "group_size": 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": 0.004361572675406933, "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": 0.004367429297417402, "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": 0.004026861861348152, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0028976250905543566, "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": 0.06141755357384682, "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": 0.06141755357384682, "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": 0.19939588010311127, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18427321314811707, "qparams": { "group_size": 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": 0.17923209071159363, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16047057509422302, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09269805997610092, "qparams": { "group_size": 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": 0.08726683259010315, "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": 0.10457221418619156, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.09645344316959381, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09437231719493866, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08195033669471741, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07692807167768478, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05306325480341911, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04600205644965172, "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": 0.044306714087724686, "qparams": { "group_size": 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": 0.04389781504869461, "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": 0.02647821232676506, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.022584082558751106, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022459978237748146, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.020381242036819458, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020116521045565605, "qparams": { "group_size": 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": 0.013647577725350857, "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": 0.013424619100987911, "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": 0.013028048910200596, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008417520672082901, "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": 0.10457221418619156, "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": 0.10457221418619156, "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": 0.2052777111530304, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17769373953342438, "qparams": { "group_size": 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": 0.16655533015727997, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14194926619529724, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09438221156597137, "qparams": { "group_size": 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": 0.08328433334827423, "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": 0.11454446613788605, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10456690192222595, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09877810627222061, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07720232754945755, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07147401571273804, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.059270910918712616, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0511065348982811, "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": 0.046452708542346954, "qparams": { "group_size": 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": 0.04532557353377342, "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": 0.030051585286855698, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02578146941959858, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025426402688026428, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02267119660973549, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021967150270938873, "qparams": { "group_size": 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": 0.017017720267176628, "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": 0.01852544955909252, "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": 0.015606371685862541, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014660780318081379, "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": 0.11454446613788605, "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": 0.11454446613788605, "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": 0.20848533511161804, "qparams": { "group_size": 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": 0.1957283467054367, "qparams": { "group_size": 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": 0.19190093874931335, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.17385633289813995, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09752210229635239, "qparams": { "group_size": 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": 0.09317931532859802, "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": 0.10780791193246841, "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": 0.0998559519648552, "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": 0.09877657145261765, "qparams": { "group_size": 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": 0.08784960955381393, "qparams": { "group_size": 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": 0.08309289813041687, "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": 0.05479395017027855, "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": 0.047715723514556885, "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": 0.046655382961034775, "qparams": { "group_size": 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": 0.046414270997047424, "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": 0.02736433781683445, "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": 0.0240040086209774, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023937370628118515, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.022053206339478493, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02189793437719345, "qparams": { "group_size": 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": 0.01433592475950718, "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": 0.014467800036072731, "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": 0.01398746483027935, "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": 0.009631989523768425, "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": 0.10780791193246841, "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": 0.10780791193246841, "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": 0.2552509009838104, "qparams": { "group_size": 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": 0.2397308498620987, "qparams": { "group_size": 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": 0.23523345589637756, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21316158771514893, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11942339688539505, "qparams": { "group_size": 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": 0.11415969580411911, "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": 0.13195806741714478, "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": 0.12228722870349884, "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": 0.12103208899497986, "qparams": { "group_size": 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": 0.10764967650175095, "qparams": { "group_size": 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": 0.10185720026493073, "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": 0.0669400691986084, "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": 0.05830162763595581, "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": 0.05703423172235489, "qparams": { "group_size": 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": 0.056728068739175797, "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": 0.033390842378139496, "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": 0.02897980809211731, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028904041275382042, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02654978632926941, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02636154368519783, "qparams": { "group_size": 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": 0.01721440814435482, "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": 0.01691669039428234, "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": 0.016789481043815613, "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": 0.010509892366826534, "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": 0.12228722870349884, "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": 0.12103208899497986, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "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": 0.25463730096817017, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.22978752851486206, "qparams": { "group_size": 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": 0.22113744914531708, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1957598626613617, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11682732403278351, "qparams": { "group_size": 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": 0.10777927935123444, "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": 0.13424387574195862, "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": 0.12379670143127441, "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": 0.12007808685302734, "qparams": { "group_size": 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": 0.10125834494829178, "qparams": { "group_size": 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": 0.09460822492837906, "qparams": { "group_size": 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": 0.06829573214054108, "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": 0.05919326841831207, "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": 0.05606172978878021, "qparams": { "group_size": 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": 0.05531317740678787, "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": 0.03422103449702263, "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": 0.02919050119817257, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029063761234283447, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026129864156246185, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025653177872300148, "qparams": { "group_size": 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": 0.01829678565263748, "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": 0.018414799124002457, "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": 0.01729622296988964, "qparams": { "group_size": 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": 0.012682385742664337, "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": 0.12007808685302734, "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": 0.12007808685302734, "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": 0.07780280709266663, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07092667371034622, "qparams": { "group_size": 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": 0.06787628680467606, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06021096929907799, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.035636670887470245, "qparams": { "group_size": 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": 0.032865628600120544, "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": 0.04173114895820618, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03857255354523659, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03645537421107292, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.031110862269997597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02922799438238144, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02116077020764351, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.018426669761538506, "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": 0.017131170257925987, "qparams": { "group_size": 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": 0.01682158000767231, "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": 0.010587926022708416, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.00894759688526392, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.008817089721560478, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.00805769395083189, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.007857048884034157, "qparams": { "group_size": 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": 0.005584829021245241, "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": 0.005740093998610973, "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": 0.005156459286808968, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003956538159400225, "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": 0.07780280709266663, "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": 0.07780280709266663, "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": 0.06723619997501373, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.061261825263500214, "qparams": { "group_size": 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": 0.058420371264219284, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.051853008568286896, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.030820943415164948, "qparams": { "group_size": 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": 0.028327353298664093, "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": 0.03657826781272888, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03358989581465721, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.031529758125543594, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.026898782700300217, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.02539079636335373, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0185257438570261, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.016023358330130577, "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": 0.014804089441895485, "qparams": { "group_size": 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": 0.014506467618048191, "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": 0.009267195127904415, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0076998150907456875, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0075751482509076595, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.00692686578258872, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.006737805902957916, "qparams": { "group_size": 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": 0.004870611242949963, "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": 0.004915356170386076, "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": 0.004462436772882938, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.003326751757413149, "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": 0.06723619997501373, "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": 0.06723619997501373, "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": 0.2099955976009369, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19403478503227234, "qparams": { "group_size": 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": 0.18854878842830658, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16868580877780914, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0978124588727951, "qparams": { "group_size": 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": 0.09197241812944412, "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": 0.11066089570522308, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10196583718061447, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09960516542196274, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08631178736686707, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08102978020906448, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05620063841342926, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0486561618745327, "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": 0.04678928107023239, "qparams": { "group_size": 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": 0.046340055763721466, "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": 0.028053155168890953, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.023860570043325424, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02371235005557537, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021503198891878128, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021214459091424942, "qparams": { "group_size": 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": 0.014485816471278667, "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": 0.014209693297743797, "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": 0.013789298012852669, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.008923675864934921, "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": 0.11066089570522308, "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": 0.11066089570522308, "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": 0.20669202506542206, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18023273348808289, "qparams": { "group_size": 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": 0.17060525715351105, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1380573958158493, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09589646011590958, "qparams": { "group_size": 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": 0.08553515374660492, "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": 0.11229778826236725, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10323342680931091, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0989912748336792, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07668737322092056, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06657470017671585, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057995639741420746, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05030638352036476, "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": 0.047008633613586426, "qparams": { "group_size": 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": 0.0462043397128582, "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": 0.029408879578113556, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025837797671556473, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025584381073713303, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.022349584847688675, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02183653600513935, "qparams": { "group_size": 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": 0.016567392274737358, "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": 0.018125392496585846, "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": 0.015576391480863094, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014295556582510471, "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": 0.11229778826236725, "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": 0.11229778826236725, "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": 0.1993962973356247, "qparams": { "group_size": 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": 0.1868104338645935, "qparams": { "group_size": 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": 0.18295221030712128, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.16550122201442719, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09352079778909683, "qparams": { "group_size": 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": 0.0891740471124649, "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": 0.10372395813465118, "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": 0.0960020199418068, "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": 0.09483157098293304, "qparams": { "group_size": 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": 0.08398372679948807, "qparams": { "group_size": 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": 0.07942073047161102, "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": 0.0528593435883522, "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": 0.04598819091916084, "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": 0.04487319663167, "qparams": { "group_size": 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": 0.0446198545396328, "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": 0.026426656171679497, "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": 0.02330821007490158, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023238657042384148, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02141575515270233, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021246690303087234, "qparams": { "group_size": 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": 0.014029065147042274, "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": 0.014396913349628448, "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": 0.013662309385836124, "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": 0.00996515154838562, "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": 0.10372395813465118, "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": 0.10372395813465118, "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": 0.2620585262775421, "qparams": { "group_size": 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": 0.2457605004310608, "qparams": { "group_size": 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": 0.24080723524093628, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21796007454395294, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12290430814027786, "qparams": { "group_size": 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": 0.11726019531488419, "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": 0.13610351085662842, "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": 0.12605327367782593, "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": 0.12462285161018372, "qparams": { "group_size": 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": 0.11046439409255981, "qparams": { "group_size": 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": 0.10439204424619675, "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": 0.06914639472961426, "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": 0.060156360268592834, "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": 0.05873555690050125, "qparams": { "group_size": 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": 0.0583941712975502, "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": 0.03451478108763695, "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": 0.029850460588932037, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029759779572486877, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027277493849396706, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02706313133239746, "qparams": { "group_size": 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": 0.017837930470705032, "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": 0.01744234561920166, "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": 0.0173561442643404, "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": 0.010815383866429329, "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": 0.11046439409255981, "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": 0.11046439409255981, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.2571476101875305, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23257794976234436, "qparams": { "group_size": 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": 0.2240303009748459, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19810593128204346, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11828285455703735, "qparams": { "group_size": 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": 0.10930265486240387, "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": 0.1356009989976883, "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": 0.1250835657119751, "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": 0.12144726514816284, "qparams": { "group_size": 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": 0.10258184373378754, "qparams": { "group_size": 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": 0.09566539525985718, "qparams": { "group_size": 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": 0.06910798698663712, "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": 0.05990378558635712, "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": 0.056833814829587936, "qparams": { "group_size": 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": 0.05610332638025284, "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": 0.034667231142520905, "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": 0.02966025471687317, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029540685936808586, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026578672230243683, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02611396089196205, "qparams": { "group_size": 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": 0.018653828650712967, "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": 0.018788283690810204, "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": 0.01768900454044342, "qparams": { "group_size": 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": 0.013073026202619076, "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": 0.12144726514816284, "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": 0.12144726514816284, "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": 0.0946882963180542, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0872935801744461, "qparams": { "group_size": 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": 0.08423132449388504, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07522957772016525, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04384457692503929, "qparams": { "group_size": 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": 0.04089631140232086, "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": 0.05050123110413551, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04664808511734009, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0446854904294014, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03867566958069801, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0363677553832531, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.025627242401242256, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022315001115202904, "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": 0.021074529737234116, "qparams": { "group_size": 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": 0.020780183374881744, "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": 0.012830597348511219, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010988756082952023, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010870720259845257, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.009966282173991203, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009779893793165684, "qparams": { "group_size": 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": 0.0067739724181592464, "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": 0.006962468847632408, "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": 0.006353924982249737, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004813358187675476, "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": 0.0946882963180542, "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": 0.0946882963180542, "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": 0.07839813828468323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07230214774608612, "qparams": { "group_size": 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": 0.06958706676959991, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06221511960029602, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03628665208816528, "qparams": { "group_size": 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": 0.03379146009683609, "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": 0.04232997074723244, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03885727748274803, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03697139769792557, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03205704689025879, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.030322259292006493, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02146853506565094, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.018559668213129044, "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": 0.017415208742022514, "qparams": { "group_size": 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": 0.01714136451482773, "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": 0.010745033621788025, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009015572257339954, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.008900310844182968, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.008167024701833725, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.007993173785507679, "qparams": { "group_size": 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": 0.0056355237029492855, "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": 0.005644300952553749, "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": 0.005251340568065643, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0037753412034362555, "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": 0.07839813828468323, "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": 0.07839813828468323, "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": 0.22613245248794556, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20992736518383026, "qparams": { "group_size": 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": 0.20444686710834503, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18348278105258942, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10573266446590424, "qparams": { "group_size": 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": 0.09981994330883026, "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": 0.11908168345689774, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10976781696081161, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10747663676738739, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0937841460108757, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0881606787443161, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06045573204755783, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.052366528660058975, "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": 0.05054287612438202, "qparams": { "group_size": 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": 0.050105493515729904, "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": 0.030169233679771423, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02571774087846279, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025571325793862343, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02325347624719143, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02297057956457138, "qparams": { "group_size": 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": 0.015511668287217617, "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": 0.015171157196164131, "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": 0.014822354540228844, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00939889345318079, "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": 0.11908168345689774, "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": 0.11908168345689774, "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": 0.2037455439567566, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17623281478881836, "qparams": { "group_size": 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": 0.16528621315956116, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.13984331488609314, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09391918778419495, "qparams": { "group_size": 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": 0.08260605484247208, "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": 0.11304431408643723, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10323497653007507, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09755692631006241, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07609851658344269, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06918046623468399, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.058148931711912155, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.050160981714725494, "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": 0.04592620208859444, "qparams": { "group_size": 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": 0.0448915995657444, "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": 0.029300019145011902, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025043079629540443, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.024690406396985054, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02173572964966297, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021081484854221344, "qparams": { "group_size": 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": 0.01634189672768116, "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": 0.017495546489953995, "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": 0.015055612660944462, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013485409319400787, "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": 0.11304431408643723, "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": 0.11304431408643723, "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": 0.19462154805660248, "qparams": { "group_size": 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": 0.18213456869125366, "qparams": { "group_size": 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": 0.17830204963684082, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.16113708913326263, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09136184304952621, "qparams": { "group_size": 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": 0.08700951188802719, "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": 0.10142471641302109, "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": 0.0939488410949707, "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": 0.09271296858787537, "qparams": { "group_size": 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": 0.08192817121744156, "qparams": { "group_size": 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": 0.07739768922328949, "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": 0.051632627844810486, "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": 0.04493971914052963, "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": 0.0437798798084259, "qparams": { "group_size": 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": 0.04350883141160011, "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": 0.025794439017772675, "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": 0.022523168474435806, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022445663809776306, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02061106637120247, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02043692022562027, "qparams": { "group_size": 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": 0.013506557792425156, "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": 0.013605250045657158, "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": 0.013122362084686756, "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": 0.009019462391734123, "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": 0.10142471641302109, "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": 0.10142471641302109, "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": 0.2645726799964905, "qparams": { "group_size": 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": 0.2478688359260559, "qparams": { "group_size": 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": 0.24276317656040192, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21965382993221283, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12430162727832794, "qparams": { "group_size": 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": 0.11846054345369339, "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": 0.13788990676403046, "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": 0.1275833547115326, "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": 0.1260737031698227, "qparams": { "group_size": 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": 0.11155688762664795, "qparams": { "group_size": 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": 0.10541960597038269, "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": 0.07013934850692749, "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": 0.06094761937856674, "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": 0.05945134907960892, "qparams": { "group_size": 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": 0.05909178406000137, "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": 0.03501669317483902, "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": 0.03026091866195202, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03016500733792782, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027629798278212547, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027397381141781807, "qparams": { "group_size": 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": 0.018186692148447037, "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": 0.01776791363954544, "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": 0.017678266391158104, "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": 0.01112247072160244, "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": 0.11155688762664795, "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": 0.11155688762664795, "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": 0.2585935890674591, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23374387621879578, "qparams": { "group_size": 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": 0.22500884532928467, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19938868284225464, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11899398267269135, "qparams": { "group_size": 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": 0.10990715026855469, "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": 0.13691528141498566, "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": 0.12610676884651184, "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": 0.122280053794384, "qparams": { "group_size": 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": 0.10326928645372391, "qparams": { "group_size": 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": 0.09646103531122208, "qparams": { "group_size": 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": 0.06966632604598999, "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": 0.06038316711783409, "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": 0.057207316160202026, "qparams": { "group_size": 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": 0.056439731270074844, "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": 0.03493397310376167, "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": 0.029862798750400543, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02972010336816311, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02676841989159584, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026279965415596962, "qparams": { "group_size": 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": 0.018753279000520706, "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": 0.018940584734082222, "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": 0.017736095935106277, "qparams": { "group_size": 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": 0.013162491843104362, "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": 0.122280053794384, "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": 0.122280053794384, "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": 0.09274926036596298, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08562713116407394, "qparams": { "group_size": 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": 0.08261829614639282, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0738661140203476, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04296484962105751, "qparams": { "group_size": 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": 0.040097516030073166, "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": 0.049604594707489014, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.045760273933410645, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04376301169395447, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03794268146157265, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03576549515128136, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02517375536262989, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.021865082904696465, "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": 0.020613137632608414, "qparams": { "group_size": 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": 0.02031009830534458, "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": 0.012587475590407848, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010664376430213451, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010538640432059765, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.009656173177063465, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009466978721320629, "qparams": { "group_size": 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": 0.006593029014766216, "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": 0.006649528630077839, "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": 0.006164254155009985, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004443598911166191, "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": 0.09274926036596298, "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": 0.09274926036596298, "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": 0.07820231467485428, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07225050032138824, "qparams": { "group_size": 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": 0.06947116553783417, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06215005740523338, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03622005507349968, "qparams": { "group_size": 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": 0.033694177865982056, "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": 0.04224901273846626, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.03898385912179947, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.03688083961606026, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03203978016972542, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.030278930440545082, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.021408051252365112, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.01860605552792549, "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": 0.017373891547322273, "qparams": { "group_size": 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": 0.017072616145014763, "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": 0.01070435531437397, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.008975483477115631, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.008846310898661613, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.008138568140566349, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.00794665701687336, "qparams": { "group_size": 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": 0.005589621607214212, "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": 0.005613876506686211, "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": 0.005180429667234421, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0037155551835894585, "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": 0.07820231467485428, "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": 0.07820231467485428, "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": 0.22082853317260742, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20513884723186493, "qparams": { "group_size": 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": 0.1997729390859604, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.17941345274448395, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10326656699180603, "qparams": { "group_size": 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": 0.09750894457101822, "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": 0.11680316925048828, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10733706504106522, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1050032526254654, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09165289998054504, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08634113520383835, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.059395384043455124, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05122759938240051, "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": 0.04938399791717529, "qparams": { "group_size": 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": 0.04894372075796127, "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": 0.0296260304749012, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025133294984698296, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02499263733625412, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.022738702595233917, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022448740899562836, "qparams": { "group_size": 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": 0.015249558724462986, "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": 0.014863298274576664, "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": 0.014510026201605797, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009236674755811691, "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": 0.11680316925048828, "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": 0.11680316925048828, "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": 0.216698557138443, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18880721926689148, "qparams": { "group_size": 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": 0.1784537434577942, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15204522013664246, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09977725148200989, "qparams": { "group_size": 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": 0.08871010690927505, "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": 0.1186022162437439, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10836714506149292, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1036662757396698, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0821666270494461, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07467782497406006, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06126735359430313, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05290321260690689, "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": 0.04901515692472458, "qparams": { "group_size": 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": 0.04807376489043236, "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": 0.03102327696979046, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.027103157714009285, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026820458471775055, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02391236461699009, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023328294977545738, "qparams": { "group_size": 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": 0.01763693243265152, "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": 0.019225239753723145, "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": 0.016467677429318428, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015256395563483238, "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": 0.1186022162437439, "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": 0.1186022162437439, "qparams": { "group_size": 128, "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": 0.17824681103229523, "qparams": { "group_size": 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": 0.16680273413658142, "qparams": { "group_size": 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": 0.16319692134857178, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1475372016429901, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08374454826116562, "qparams": { "group_size": 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": 0.07968304306268692, "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": 0.09313883632421494, "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": 0.08624210208654404, "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": 0.08494754135608673, "qparams": { "group_size": 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": 0.0750720202922821, "qparams": { "group_size": 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": 0.07095041871070862, "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": 0.04740641638636589, "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": 0.04129646718502045, "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": 0.0401727594435215, "qparams": { "group_size": 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": 0.03990703821182251, "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": 0.023700537160038948, "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": 0.020754234865307808, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02067510038614273, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019002940505743027, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01883915811777115, "qparams": { "group_size": 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": 0.012456389144062996, "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": 0.012664346024394035, "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": 0.012083999812602997, "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": 0.008552559651434422, "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": 0.09313883632421494, "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": 0.09313883632421494, "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": 0.25618070363998413, "qparams": { "group_size": 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": 0.24005205929279327, "qparams": { "group_size": 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": 0.23505105078220367, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21279728412628174, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12067094445228577, "qparams": { "group_size": 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": 0.11497513949871063, "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": 0.1340310275554657, "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": 0.12401427328586578, "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": 0.12241710722446442, "qparams": { "group_size": 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": 0.10831786692142487, "qparams": { "group_size": 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": 0.10244134813547134, "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": 0.06833764165639877, "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": 0.05937473475933075, "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": 0.05787087604403496, "qparams": { "group_size": 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": 0.05751561000943184, "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": 0.03412038832902908, "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": 0.0298458244651556, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029748618602752686, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027344217523932457, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027121029794216156, "qparams": { "group_size": 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": 0.017927952110767365, "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": 0.01813451573252678, "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": 0.017417103052139282, "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": 0.01217395905405283, "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": 0.12241710722446442, "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": 0.12067094445228577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.2445208877325058, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.22146570682525635, "qparams": { "group_size": 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": 0.21346968412399292, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1891525834798813, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11258211731910706, "qparams": { "group_size": 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": 0.10421182960271835, "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": 0.13009214401245117, "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": 0.11902950704097748, "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": 0.1155298724770546, "qparams": { "group_size": 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": 0.09786161035299301, "qparams": { "group_size": 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": 0.09149106591939926, "qparams": { "group_size": 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": 0.06612856686115265, "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": 0.05715465173125267, "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": 0.05423036590218544, "qparams": { "group_size": 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": 0.05353587120771408, "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": 0.03330833092331886, "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": 0.02858573943376541, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02845335192978382, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02573978528380394, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025297200307250023, "qparams": { "group_size": 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": 0.018258119001984596, "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": 0.018474752083420753, "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": 0.017328381538391113, "qparams": { "group_size": 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": 0.01327047124505043, "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": 0.11902950704097748, "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": 0.11902950704097748, "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": 0.10243844240903854, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09474404901266098, "qparams": { "group_size": 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": 0.09149548411369324, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08192159235477448, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.047471027821302414, "qparams": { "group_size": 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": 0.044420499354600906, "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": 0.05467243492603302, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05049356445670128, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04833917319774628, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.042033758014440536, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.039650071412324905, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027736574411392212, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.024111270904541016, "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": 0.022777464240789413, "qparams": { "group_size": 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": 0.022456739097833633, "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": 0.01387548353523016, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011783162131905556, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011653173714876175, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0106895063072443, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010487383231520653, "qparams": { "group_size": 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": 0.007272791117429733, "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": 0.007339141797274351, "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": 0.006824157666414976, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0049070636741817, "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": 0.10243844240903854, "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": 0.10243844240903854, "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": 0.08425169438123703, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.07795918732881546, "qparams": { "group_size": 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": 0.0750063955783844, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.06719367951154709, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.03903000429272652, "qparams": { "group_size": 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": 0.036382805556058884, "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": 0.04537297040224075, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04192015901207924, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.039727553725242615, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.034581273794174194, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.032692208886146545, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02298746071755886, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.020008506253361702, "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": 0.018726501613855362, "qparams": { "group_size": 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": 0.01841532066464424, "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": 0.01150161400437355, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.009677371010184288, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009542914107441902, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.008784723468124866, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.008586470037698746, "qparams": { "group_size": 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": 0.00602398905903101, "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": 0.006051507778465748, "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": 0.005611864849925041, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004015998914837837, "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": 0.08425169438123703, "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": 0.08425169438123703, "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": 0.2328575998544693, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21655257046222687, "qparams": { "group_size": 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": 0.2111092358827591, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18963947892189026, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10902564227581024, "qparams": { "group_size": 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": 0.10299208015203476, "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": 0.12358903884887695, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11317116022109985, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11075979471206665, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09682340174913406, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09147033095359802, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06278793513774872, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.054018862545490265, "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": 0.052119530737400055, "qparams": { "group_size": 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": 0.05166792869567871, "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": 0.03134671598672867, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02650991827249527, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026358185335993767, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.024000151082873344, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02370447665452957, "qparams": { "group_size": 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": 0.01614086702466011, "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": 0.015654506161808968, "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": 0.015341660007834435, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009697114117443562, "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": 0.11317116022109985, "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": 0.11317116022109985, "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": 0.20748113095760345, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18384882807731628, "qparams": { "group_size": 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": 0.1757718324661255, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14967817068099976, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09630164504051208, "qparams": { "group_size": 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": 0.08775801956653595, "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": 0.11160381883382797, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10228335857391357, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09900660812854767, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07977772504091263, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0718090832233429, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.057760369032621384, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04988078400492668, "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": 0.04720405489206314, "qparams": { "group_size": 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": 0.04657474905252457, "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": 0.029388202354311943, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025951679795980453, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025754177942872047, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02293107658624649, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022518672049045563, "qparams": { "group_size": 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": 0.016945993527770042, "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": 0.018095985054969788, "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": 0.01616525650024414, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014336825348436832, "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": 0.11160381883382797, "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": 0.11160381883382797, "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": 0.18001322448253632, "qparams": { "group_size": 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": 0.16858869791030884, "qparams": { "group_size": 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": 0.16487327218055725, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1492023915052414, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08472573757171631, "qparams": { "group_size": 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": 0.08063056319952011, "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": 0.09431330114603043, "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": 0.08728138357400894, "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": 0.08596320450305939, "qparams": { "group_size": 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": 0.07600875943899155, "qparams": { "group_size": 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": 0.07195588946342468, "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": 0.04818986728787422, "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": 0.04188855364918709, "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": 0.04074045643210411, "qparams": { "group_size": 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": 0.0404696986079216, "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": 0.024087490513920784, "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": 0.021235203370451927, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021156925708055496, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019506104290485382, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.019333967939019203, "qparams": { "group_size": 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": 0.012823785655200481, "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": 0.01324926782399416, "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": 0.012447860091924667, "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": 0.009293022565543652, "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": 0.09431330114603043, "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": 0.09431330114603043, "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": 0.2534736692905426, "qparams": { "group_size": 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": 0.23756468296051025, "qparams": { "group_size": 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": 0.23257233202457428, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2106579691171646, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11917778849601746, "qparams": { "group_size": 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": 0.11352527886629105, "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": 0.13255350291728973, "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": 0.12255725264549255, "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": 0.12086737900972366, "qparams": { "group_size": 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": 0.10698743164539337, "qparams": { "group_size": 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": 0.10119219124317169, "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": 0.06743521243333817, "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": 0.05854027345776558, "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": 0.057014741003513336, "qparams": { "group_size": 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": 0.05665118247270584, "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": 0.033660538494586945, "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": 0.02906668372452259, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02895735576748848, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026552600786089897, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02631762996315956, "qparams": { "group_size": 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": 0.017470832914114, "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": 0.017140671610832214, "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": 0.016938164830207825, "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": 0.010797032155096531, "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": 0.12255725264549255, "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": 0.12086737900972366, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "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": 0.25291723012924194, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.22947464883327484, "qparams": { "group_size": 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": 0.22115179896354675, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19607976078987122, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11661169677972794, "qparams": { "group_size": 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": 0.10792788863182068, "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": 0.13413873314857483, "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": 0.12340562045574188, "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": 0.11966615170240402, "qparams": { "group_size": 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": 0.10150281339883804, "qparams": { "group_size": 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": 0.09478338062763214, "qparams": { "group_size": 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": 0.06831721216440201, "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": 0.05922670662403107, "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": 0.056190524250268936, "qparams": { "group_size": 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": 0.055449388921260834, "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": 0.034331563860177994, "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": 0.02956194430589676, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029423289000988007, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026620684191584587, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02616049163043499, "qparams": { "group_size": 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": 0.0185585655272007, "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": 0.019044915214180946, "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": 0.01757015474140644, "qparams": { "group_size": 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": 0.013619789853692055, "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": 0.11966615170240402, "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": 0.11966615170240402, "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": 0.10856654495000839, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10055641084909439, "qparams": { "group_size": 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": 0.09720080345869064, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08715330064296722, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0505044125020504, "qparams": { "group_size": 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": 0.047284919768571854, "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": 0.05808871239423752, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05364043638110161, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.051366161555051804, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04480500519275665, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04234887287020683, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029603872448205948, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02576475962996483, "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": 0.02436015196144581, "qparams": { "group_size": 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": 0.024019641801714897, "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": 0.014842587523162365, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012843894772231579, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012705379165709019, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011734293773770332, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011526932939887047, "qparams": { "group_size": 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": 0.007923568598926067, "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": 0.008321360684931278, "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": 0.00745637621730566, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005979232490062714, "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": 0.10856654495000839, "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": 0.10856654495000839, "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": 0.09263448417186737, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08593566715717316, "qparams": { "group_size": 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": 0.08282690495252609, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0743415430188179, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04304906725883484, "qparams": { "group_size": 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": 0.04025101289153099, "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": 0.04991701990365982, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04605705291032791, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.043814342468976974, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.038233496248722076, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03621507063508034, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.025344468653202057, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022008029744029045, "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": 0.020670069381594658, "qparams": { "group_size": 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": 0.02034698612987995, "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": 0.012680239975452423, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01070157065987587, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010564194060862064, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.009737190790474415, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009532364085316658, "qparams": { "group_size": 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": 0.006662378553301096, "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": 0.006699639838188887, "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": 0.006223213858902454, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004483999218791723, "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": 0.09263448417186737, "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": 0.09263448417186737, "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": 0.2275456190109253, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2117716670036316, "qparams": { "group_size": 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": 0.2063957154750824, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18565283715724945, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10649152845144272, "qparams": { "group_size": 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": 0.10067950934171677, "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": 0.12085999548435211, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11066722869873047, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10821894556283951, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09479755163192749, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08956463634967804, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06144581735134125, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05285602807998657, "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": 0.05095435678958893, "qparams": { "group_size": 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": 0.050491318106651306, "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": 0.030656680464744568, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02598850056529045, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025831760838627815, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023578204214572906, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023286793380975723, "qparams": { "group_size": 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": 0.01585252955555916, "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": 0.015453696250915527, "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": 0.01507845800369978, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00972895510494709, "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": 0.12085999548435211, "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": 0.12085999548435211, "qparams": { "group_size": 128, "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": 0.22140511870384216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1943594217300415, "qparams": { "group_size": 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": 0.18431277573108673, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15171237289905548, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10249464958906174, "qparams": { "group_size": 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": 0.09184696525335312, "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": 0.12167788296937943, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11124783754348755, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10571926832199097, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08355212956666946, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0743585005402565, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06274186074733734, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05453028157353401, "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": 0.05060912296175957, "qparams": { "group_size": 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": 0.049671340733766556, "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": 0.031990692019462585, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02840171940624714, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028107021003961563, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.025054223835468292, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02447216957807541, "qparams": { "group_size": 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": 0.018501847982406616, "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": 0.02056860737502575, "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": 0.017353560775518417, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01670684479176998, "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": 0.12167788296937943, "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": 0.12167788296937943, "qparams": { "group_size": 128, "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": 0.18416818976402283, "qparams": { "group_size": 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": 0.17246636748313904, "qparams": { "group_size": 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": 0.1686607152223587, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.15257592499256134, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08669758588075638, "qparams": { "group_size": 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": 0.0824790894985199, "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": 0.09650419652462006, "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": 0.08935991674661636, "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": 0.08793920278549194, "qparams": { "group_size": 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": 0.07775940001010895, "qparams": { "group_size": 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": 0.07360758632421494, "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": 0.04926110431551933, "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": 0.042884353548288345, "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": 0.04168405383825302, "qparams": { "group_size": 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": 0.041398242115974426, "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": 0.02463178150355816, "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": 0.02171439491212368, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021631132811307907, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019942672923207283, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01976800337433815, "qparams": { "group_size": 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": 0.01307048462331295, "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": 0.013531964272260666, "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": 0.012675454840064049, "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": 0.009482242166996002, "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": 0.09650419652462006, "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": 0.09650419652462006, "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": 0.2551117539405823, "qparams": { "group_size": 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": 0.23911316692829132, "qparams": { "group_size": 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": 0.23420336842536926, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21202534437179565, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11999358981847763, "qparams": { "group_size": 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": 0.11426601558923721, "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": 0.1335049867630005, "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": 0.12342681735754013, "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": 0.12168648093938828, "qparams": { "group_size": 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": 0.10775692015886307, "qparams": { "group_size": 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": 0.10190609842538834, "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": 0.06782716512680054, "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": 0.0589696541428566, "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": 0.0574091374874115, "qparams": { "group_size": 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": 0.05703921616077423, "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": 0.03387666866183281, "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": 0.029253220185637474, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0291427168995142, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026726670563220978, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026485305279493332, "qparams": { "group_size": 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": 0.017557691782712936, "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": 0.01723993755877018, "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": 0.017015602439641953, "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": 0.010830136947333813, "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": 0.12168648093938828, "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": 0.11999358981847763, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.25780725479125977, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23375266790390015, "qparams": { "group_size": 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": 0.22519782185554504, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19940771162509918, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1190289780497551, "qparams": { "group_size": 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": 0.11004873365163803, "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": 0.13701260089874268, "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": 0.12613892555236816, "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": 0.12217291444540024, "qparams": { "group_size": 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": 0.103441022336483, "qparams": { "group_size": 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": 0.09662342816591263, "qparams": { "group_size": 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": 0.0699409618973732, "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": 0.060631148517131805, "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": 0.0573936402797699, "qparams": { "group_size": 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": 0.056631434708833694, "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": 0.03515422344207764, "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": 0.030314896255731583, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030164450407028198, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02729889005422592, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026820851489901543, "qparams": { "group_size": 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": 0.019151385873556137, "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": 0.019690459594130516, "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": 0.018135204911231995, "qparams": { "group_size": 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": 0.014233001507818699, "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": 0.12217291444540024, "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": 0.12217291444540024, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11357627063989639, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10530797392129898, "qparams": { "group_size": 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": 0.1018337681889534, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09134479612112045, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0527956523001194, "qparams": { "group_size": 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": 0.049462273716926575, "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": 0.060573454946279526, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05596313625574112, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05369022116065025, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04686747118830681, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04420166090130806, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03079947456717491, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.026786385104060173, "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": 0.025359883904457092, "qparams": { "group_size": 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": 0.02502017840743065, "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": 0.015407822094857693, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013201158493757248, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013057364150881767, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012024689465761185, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011807331815361977, "qparams": { "group_size": 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": 0.008123330771923065, "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": 0.008323455229401588, "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": 0.007644040044397116, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005711155012249947, "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": 0.11357627063989639, "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": 0.11357627063989639, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.0932774767279625, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08652622252702713, "qparams": { "group_size": 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": 0.08335167914628983, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0747782364487648, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04324667155742645, "qparams": { "group_size": 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": 0.04039020091295242, "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": 0.0502239391207695, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.046411678194999695, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.043992139399051666, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03841068223118782, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03634429723024368, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02543947473168373, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02215660735964775, "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": 0.020757928490638733, "qparams": { "group_size": 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": 0.020417289808392525, "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": 0.012732929550111294, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01074804738163948, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010599170811474323, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.009782476350665092, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009565741755068302, "qparams": { "group_size": 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": 0.006681892089545727, "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": 0.006740220356732607, "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": 0.006235891953110695, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004499003291130066, "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": 0.0932774767279625, "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": 0.0932774767279625, "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": 0.22688430547714233, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2111119031906128, "qparams": { "group_size": 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": 0.20530222356319427, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18461748957633972, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10610590130090714, "qparams": { "group_size": 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": 0.10009937733411789, "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": 0.12040242552757263, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11089058965444565, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10781615972518921, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.094415083527565, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08895634114742279, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06116238981485367, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0529707670211792, "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": 0.050779011100530624, "qparams": { "group_size": 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": 0.050250012427568436, "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": 0.030513789504766464, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025937680155038834, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02573465369641781, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023533407598733902, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0231940858066082, "qparams": { "group_size": 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": 0.015797194093465805, "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": 0.015527188777923584, "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": 0.015000739134848118, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009811274707317352, "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": 0.12040242552757263, "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": 0.12040242552757263, "qparams": { "group_size": 128, "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": 0.2303205132484436, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20523357391357422, "qparams": { "group_size": 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": 0.19595253467559814, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1669521927833557, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10694672167301178, "qparams": { "group_size": 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": 0.09716027230024338, "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": 0.12546606361865997, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1152595654129982, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11021900177001953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09002067148685455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08094578981399536, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06469225883483887, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05636210739612579, "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": 0.05259527266025543, "qparams": { "group_size": 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": 0.051701076328754425, "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": 0.03281114995479584, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.029140744358301163, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02882886677980423, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026065479964017868, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025493208318948746, "qparams": { "group_size": 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": 0.018636438995599747, "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": 0.020683621987700462, "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": 0.017522068694233894, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.016499610617756844, "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": 0.1152595654129982, "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": 0.1152595654129982, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.17683620750904083, "qparams": { "group_size": 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": 0.16557373106479645, "qparams": { "group_size": 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": 0.16185009479522705, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.14645777642726898, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08322448283433914, "qparams": { "group_size": 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": 0.07916346192359924, "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": 0.09278014302253723, "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": 0.08584467321634293, "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": 0.0844372883439064, "qparams": { "group_size": 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": 0.07466249912977219, "qparams": { "group_size": 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": 0.07063613831996918, "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": 0.04743286222219467, "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": 0.04125085100531578, "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": 0.04006408527493477, "qparams": { "group_size": 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": 0.03978760167956352, "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": 0.023731984198093414, "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": 0.020988980308175087, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020908895879983902, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01930108666419983, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01912970095872879, "qparams": { "group_size": 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": 0.012712787836790085, "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": 0.013249460607767105, "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": 0.012330913916230202, "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": 0.009458072483539581, "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": 0.09278014302253723, "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": 0.09278014302253723, "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": 0.24932187795639038, "qparams": { "group_size": 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": 0.23348918557167053, "qparams": { "group_size": 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": 0.22849871218204498, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2068280130624771, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11710670590400696, "qparams": { "group_size": 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": 0.11148812621831894, "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": 0.1303655505180359, "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": 0.12054415047168732, "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": 0.11879678070545197, "qparams": { "group_size": 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": 0.10506319999694824, "qparams": { "group_size": 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": 0.0993410050868988, "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": 0.06622644513845444, "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": 0.05755631998181343, "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": 0.0560273602604866, "qparams": { "group_size": 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": 0.0556630864739418, "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": 0.03307453170418739, "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": 0.028548693284392357, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028444822877645493, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026063762605190277, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025829074904322624, "qparams": { "group_size": 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": 0.01716729812324047, "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": 0.01682421565055847, "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": 0.016638802364468575, "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": 0.010579738765954971, "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": 0.12054415047168732, "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": 0.12054415047168732, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.2591531276702881, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23552806675434113, "qparams": { "group_size": 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": 0.22742104530334473, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20190201699733734, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1196579560637474, "qparams": { "group_size": 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": 0.11100678145885468, "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": 0.13663916289806366, "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": 0.12617622315883636, "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": 0.12270178645849228, "qparams": { "group_size": 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": 0.10435547679662704, "qparams": { "group_size": 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": 0.09753753989934921, "qparams": { "group_size": 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": 0.06978701800107956, "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": 0.06057732179760933, "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": 0.05764051899313927, "qparams": { "group_size": 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": 0.056935008615255356, "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": 0.03504583612084389, "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": 0.030321093276143074, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030191440135240555, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027339817956089973, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026894761249423027, "qparams": { "group_size": 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": 0.018993161618709564, "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": 0.01950380578637123, "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": 0.018071653321385384, "qparams": { "group_size": 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": 0.013951235450804234, "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": 0.12270178645849228, "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": 0.12270178645849228, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11182600259780884, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10375099629163742, "qparams": { "group_size": 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": 0.10029934346675873, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09000252187252045, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05195087939500809, "qparams": { "group_size": 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": 0.048679057508707047, "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": 0.05977354571223259, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05518714711070061, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05283635854721069, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04615052044391632, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04356934130191803, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.030356070026755333, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.026383616030216217, "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": 0.02492799423635006, "qparams": { "group_size": 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": 0.024577762931585312, "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": 0.015183452516794205, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012906442396342754, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012761273421347141, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011746147647500038, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011523270048201084, "qparams": { "group_size": 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": 0.007967348210513592, "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": 0.008051923476159573, "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": 0.007471325807273388, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005409712437540293, "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": 0.11182600259780884, "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": 0.11182600259780884, "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": 0.09397925436496735, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.08727134019136429, "qparams": { "group_size": 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": 0.08409518003463745, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07549649477005005, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04365959018468857, "qparams": { "group_size": 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": 0.040779199451208115, "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": 0.05063144117593765, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.046872034668922424, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04438013210892677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.03883793205022812, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.036703743040561676, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.025668587535619736, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.022377246990799904, "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": 0.020940441638231277, "qparams": { "group_size": 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": 0.02059374377131462, "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": 0.012839112430810928, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.010815151035785675, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010663474909961224, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.009847634471952915, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009628619998693466, "qparams": { "group_size": 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": 0.00672944076359272, "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": 0.006756185088306665, "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": 0.0062756105326116085, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004475586581975222, "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": 0.09397925436496735, "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": 0.09397925436496735, "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": 0.22794784605503082, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21213921904563904, "qparams": { "group_size": 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": 0.20635581016540527, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.18544460833072662, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10660376399755478, "qparams": { "group_size": 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": 0.10057643055915833, "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": 0.12119683623313904, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11147449165582657, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10832872986793518, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09485064446926117, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08943549543619156, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0615273080766201, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.053225044161081314, "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": 0.05102118104696274, "qparams": { "group_size": 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": 0.050485000014305115, "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": 0.030723970383405685, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02606900781393051, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02586938440799713, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02365497499704361, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023315098136663437, "qparams": { "group_size": 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": 0.01589100807905197, "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": 0.015618682838976383, "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": 0.015070979483425617, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.009885821491479874, "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": 0.12119683623313904, "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": 0.12119683623313904, "qparams": { "group_size": 128, "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": 0.22396065294742584, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19939301908016205, "qparams": { "group_size": 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": 0.18743501603603363, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.16150306165218353, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10346806049346924, "qparams": { "group_size": 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": 0.0927460640668869, "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": 0.12660038471221924, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11632202565670013, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10786016285419464, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08792749792337418, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08111657202243805, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06519047915935516, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.056619398295879364, "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": 0.05078728497028351, "qparams": { "group_size": 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": 0.049310971051454544, "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": 0.03292395919561386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.028003888204693794, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02746571972966194, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.025166772305965424, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02426750585436821, "qparams": { "group_size": 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": 0.018290484324097633, "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": 0.020060302689671516, "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": 0.01647145301103592, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015621611848473549, "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": 0.11632202565670013, "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": 0.11632202565670013, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "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": 0.17866425216197968, "qparams": { "group_size": 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": 0.16723161935806274, "qparams": { "group_size": 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": 0.1633918136358261, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.14780917763710022, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08410070836544037, "qparams": { "group_size": 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": 0.07996899634599686, "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": 0.0939251184463501, "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": 0.08692550659179688, "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": 0.08535086363554001, "qparams": { "group_size": 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": 0.07544364035129547, "qparams": { "group_size": 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": 0.07140197604894638, "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": 0.04800841212272644, "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": 0.04180590435862541, "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": 0.04054639860987663, "qparams": { "group_size": 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": 0.04025360941886902, "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": 0.024034807458519936, "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": 0.02131596952676773, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021222008392214775, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01961379498243332, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.019428353756666183, "qparams": { "group_size": 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": 0.01288436260074377, "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": 0.013561449944972992, "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": 0.012476921081542969, "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": 0.009783831425011158, "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": 0.0939251184463501, "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": 0.0939251184463501, "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": 0.24724635481834412, "qparams": { "group_size": 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": 0.23150606453418732, "qparams": { "group_size": 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": 0.2264777421951294, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2048265039920807, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1161445826292038, "qparams": { "group_size": 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": 0.11053428053855896, "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": 0.1295081377029419, "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": 0.11964169144630432, "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": 0.11782759428024292, "qparams": { "group_size": 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": 0.10413889586925507, "qparams": { "group_size": 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": 0.09841232001781464, "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": 0.06574881076812744, "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": 0.05716363340616226, "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": 0.05559062212705612, "qparams": { "group_size": 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": 0.05522007495164871, "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": 0.032832641154527664, "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": 0.02837289683520794, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028257738798856735, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.025904297828674316, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02565910294651985, "qparams": { "group_size": 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": 0.017003806307911873, "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": 0.01680128276348114, "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": 0.016446325927972794, "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": 0.010661767795681953, "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": 0.11964169144630432, "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": 0.11964169144630432, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.26281455159187317, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23883916437625885, "qparams": { "group_size": 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": 0.2304098904132843, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20438699424266815, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12153498083353043, "qparams": { "group_size": 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": 0.11263774335384369, "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": 0.13904695212841034, "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": 0.12837564945220947, "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": 0.12464246153831482, "qparams": { "group_size": 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": 0.10583959519863129, "qparams": { "group_size": 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": 0.09880796074867249, "qparams": { "group_size": 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": 0.07108555734157562, "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": 0.06165182963013649, "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": 0.05853859707713127, "qparams": { "group_size": 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": 0.05779089406132698, "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": 0.035722434520721436, "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": 0.030782150104641914, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030653292313218117, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02772609144449234, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02726057916879654, "qparams": { "group_size": 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": 0.01939532719552517, "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": 0.019791148602962494, "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": 0.018439801409840584, "qparams": { "group_size": 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": 0.014143336564302444, "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": 0.12153498083353043, "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": 0.12153498083353043, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11740517616271973, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10893996804952621, "qparams": { "group_size": 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": 0.10543704777956009, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09459373354911804, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05459029600024223, "qparams": { "group_size": 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": 0.051208820194005966, "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": 0.06258108466863632, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05783921480178833, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05550825595855713, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04849676787853241, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.045723967254161835, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.031772300601005554, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02762688510119915, "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": 0.02617683820426464, "qparams": { "group_size": 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": 0.0258285291492939, "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": 0.015888050198554993, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01349399983882904, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01334886159747839, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01226570550352335, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012043075636029243, "qparams": { "group_size": 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": 0.008295608684420586, "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": 0.008326222188770771, "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": 0.007806941866874695, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005491999909281731, "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": 0.11740517616271973, "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": 0.11740517616271973, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.09739162772893906, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09050733596086502, "qparams": { "group_size": 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": 0.08724581450223923, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07829058915376663, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04525606334209442, "qparams": { "group_size": 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": 0.04229867830872536, "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": 0.05237792432308197, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04853453487157822, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.046028219163417816, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04027023911476135, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.038048673421144485, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026540694758296013, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02317139506340027, "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": 0.021696094423532486, "qparams": { "group_size": 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": 0.02134368009865284, "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": 0.013264266774058342, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011182073503732681, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011023206636309624, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01017716620117426, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.009949947707355022, "qparams": { "group_size": 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": 0.006933882832527161, "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": 0.006939897779375315, "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": 0.006470994092524052, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004536580294370651, "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": 0.09739162772893906, "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": 0.09739162772893906, "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": 0.2341752052307129, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21789570152759552, "qparams": { "group_size": 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": 0.21176797151565552, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.19031400978565216, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10964335501194, "qparams": { "group_size": 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": 0.10334723442792892, "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": 0.12468541413545609, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1147691011428833, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1113806962966919, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09752820432186127, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0919065922498703, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06339146196842194, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05491037666797638, "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": 0.052503202110528946, "qparams": { "group_size": 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": 0.05193686857819557, "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": 0.03168954700231552, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026899373158812523, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026680417358875275, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.024422306567430496, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0240582674741745, "qparams": { "group_size": 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": 0.01647448167204857, "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": 0.016243474557995796, "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": 0.015611141920089722, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.010409509763121605, "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": 0.1147691011428833, "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": 0.1147691011428833, "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": 0.2345999777317047, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.210809126496315, "qparams": { "group_size": 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": 0.20196634531021118, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.17517070472240448, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1092219203710556, "qparams": { "group_size": 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": 0.09986387193202972, "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": 0.1275932788848877, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11687245219945908, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11217540502548218, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0929548516869545, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08529236167669296, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06550771743059158, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0567602664232254, "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": 0.05327700823545456, "qparams": { "group_size": 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": 0.05244534835219383, "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": 0.03301765397191048, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.029031328856945038, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028750373050570488, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02600313536822796, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025485500693321228, "qparams": { "group_size": 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": 0.018418466672301292, "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": 0.020010551437735558, "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": 0.01735575683414936, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01554242055863142, "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": 0.11687245219945908, "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": 0.11687245219945908, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.18246617913246155, "qparams": { "group_size": 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": 0.17077724635601044, "qparams": { "group_size": 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": 0.16690902411937714, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.15092533826828003, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08596307784318924, "qparams": { "group_size": 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": 0.08172053098678589, "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": 0.09610427170991898, "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": 0.08883120864629745, "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": 0.0872146263718605, "qparams": { "group_size": 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": 0.07709436118602753, "qparams": { "group_size": 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": 0.07297364622354507, "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": 0.04912154749035835, "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": 0.042738188058137894, "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": 0.04145662486553192, "qparams": { "group_size": 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": 0.04116186499595642, "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": 0.02459251694381237, "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": 0.02182147651910782, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02172892726957798, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.020086213946342468, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0199052095413208, "qparams": { "group_size": 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": 0.013222260400652885, "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": 0.013925273902714252, "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": 0.012810282409191132, "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": 0.010090530849993229, "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": 0.09610427170991898, "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": 0.09610427170991898, "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": 0.24860480427742004, "qparams": { "group_size": 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": 0.23281601071357727, "qparams": { "group_size": 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": 0.22773133218288422, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20600523054599762, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11684592813253403, "qparams": { "group_size": 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": 0.11117316037416458, "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": 0.130305215716362, "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": 0.1203852966427803, "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": 0.11855164170265198, "qparams": { "group_size": 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": 0.10477428138256073, "qparams": { "group_size": 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": 0.09906158596277237, "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": 0.06629401445388794, "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": 0.05750881880521774, "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": 0.05591201409697533, "qparams": { "group_size": 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": 0.05552763491868973, "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": 0.033102236688137054, "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": 0.02848491258919239, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028374025598168373, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0259934663772583, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02575116790831089, "qparams": { "group_size": 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": 0.017170052975416183, "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": 0.016795363277196884, "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": 0.01660124585032463, "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": 0.01054953970015049, "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": 0.1203852966427803, "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": 0.1203852966427803, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.266154408454895, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24190238118171692, "qparams": { "group_size": 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": 0.23358683288097382, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20705710351467133, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12318789213895798, "qparams": { "group_size": 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": 0.11424849927425385, "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": 0.14065392315387726, "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": 0.12976308166980743, "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": 0.1262912005186081, "qparams": { "group_size": 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": 0.10716401040554047, "qparams": { "group_size": 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": 0.10012384504079819, "qparams": { "group_size": 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": 0.07194703817367554, "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": 0.06235986202955246, "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": 0.05933872610330582, "qparams": { "group_size": 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": 0.05861954018473625, "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": 0.036137644201517105, "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": 0.03122960776090622, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03110954724252224, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028119653463363647, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02766963467001915, "qparams": { "group_size": 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": 0.019627796486020088, "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": 0.02009005844593048, "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": 0.01869926042854786, "qparams": { "group_size": 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": 0.01439758576452732, "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": 0.10716401040554047, "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": 0.10716401040554047, "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": 0.1245156079530716, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11561162024736404, "qparams": { "group_size": 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": 0.11193951219320297, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10051926970481873, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05799268186092377, "qparams": { "group_size": 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": 0.05446156859397888, "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": 0.06643669307231903, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06135593727231026, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05897650867700577, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.051564496010541916, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.048708926886320114, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03375961631536484, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029398959130048752, "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": 0.027877824380993843, "qparams": { "group_size": 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": 0.027518469840288162, "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": 0.01691100187599659, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014491530135273933, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01434056181460619, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013209586031734943, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012980222702026367, "qparams": { "group_size": 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": 0.00890442170202732, "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": 0.009100803174078465, "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": 0.008403546176850796, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006227593868970871, "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": 0.11561162024736404, "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": 0.11561162024736404, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.09915882349014282, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09211462736129761, "qparams": { "group_size": 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": 0.08858457207679749, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07951749116182327, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04603005200624466, "qparams": { "group_size": 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": 0.042952775955200195, "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": 0.05359366908669472, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.049686044454574585, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.046797920018434525, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.040987562388181686, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03877756744623184, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027162402868270874, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023712223395705223, "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": 0.022087011486291885, "qparams": { "group_size": 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": 0.021692264825105667, "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": 0.01358291320502758, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011417838744819164, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011237995699048042, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.010402720421552658, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010152817703783512, "qparams": { "group_size": 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": 0.0071238018572330475, "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": 0.00715511804446578, "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": 0.006620799656957388, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004724384751170874, "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": 0.09915882349014282, "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": 0.09915882349014282, "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": 0.2432795614004135, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2265312224626541, "qparams": { "group_size": 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": 0.2203063815832138, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1981751024723053, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11396123468875885, "qparams": { "group_size": 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": 0.1075408011674881, "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": 0.1294226348400116, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11902926117181778, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11572720110416412, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.10145535320043564, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09573493897914886, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06582903861999512, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05694340169429779, "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": 0.05457223951816559, "qparams": { "group_size": 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": 0.05400148034095764, "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": 0.032904546707868576, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.027936628088355064, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027716834098100662, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.025381186977028847, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02501579560339451, "qparams": { "group_size": 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": 0.01709279790520668, "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": 0.016836754977703094, "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": 0.016210012137889862, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.010779280215501785, "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": 0.11902926117181778, "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": 0.11902926117181778, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2398482710123062, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21341755986213684, "qparams": { "group_size": 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": 0.2047438621520996, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1735217124223709, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11200173944234848, "qparams": { "group_size": 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": 0.10252712666988373, "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": 0.12907974421977997, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11818207055330276, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11479178816080093, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09308202564716339, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08356136083602905, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06662115454673767, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05768651142716408, "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": 0.05494173616170883, "qparams": { "group_size": 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": 0.0542767196893692, "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": 0.03369670361280441, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030088085681200027, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029894109815359116, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026570815593004227, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02617444097995758, "qparams": { "group_size": 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": 0.01900070160627365, "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": 0.020784612745046616, "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": 0.01819859817624092, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01641208492219448, "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": 0.11818207055330276, "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": 0.11818207055330276, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.1805800497531891, "qparams": { "group_size": 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": 0.1690908819437027, "qparams": { "group_size": 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": 0.16523897647857666, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.14951549470424652, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08514860272407532, "qparams": { "group_size": 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": 0.08095025271177292, "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": 0.09505540132522583, "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": 0.08794089406728745, "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": 0.0864008292555809, "qparams": { "group_size": 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": 0.07634200900793076, "qparams": { "group_size": 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": 0.07226747274398804, "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": 0.04859056696295738, "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": 0.04232233390212059, "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": 0.04107462242245674, "qparams": { "group_size": 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": 0.04077913984656334, "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": 0.02433783747255802, "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": 0.02159457840025425, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021509310230612755, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019873879849910736, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.019688189029693604, "qparams": { "group_size": 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": 0.013070877641439438, "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": 0.013756641186773777, "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": 0.012667475268244743, "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": 0.009944834746420383, "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": 0.09505540132522583, "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": 0.09505540132522583, "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": 0.2509404122829437, "qparams": { "group_size": 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": 0.23504161834716797, "qparams": { "group_size": 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": 0.2300230711698532, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2080896645784378, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11802875250577927, "qparams": { "group_size": 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": 0.11231023818254471, "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": 0.13146358728408813, "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": 0.121524877846241, "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": 0.1197381243109703, "qparams": { "group_size": 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": 0.1058494821190834, "qparams": { "group_size": 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": 0.10001793503761292, "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": 0.06682609021663666, "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": 0.05804789811372757, "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": 0.05646079406142235, "qparams": { "group_size": 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": 0.05609049275517464, "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": 0.03337337449193001, "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": 0.028763433918356895, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028650745749473572, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026236357167363167, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02599935233592987, "qparams": { "group_size": 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": 0.01730002462863922, "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": 0.016939671710133553, "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": 0.01674681156873703, "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": 0.0106261121109128, "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": 0.121524877846241, "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": 0.121524877846241, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.26475992798805237, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24040941894054413, "qparams": { "group_size": 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": 0.2318534255027771, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20607689023017883, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12226102501153946, "qparams": { "group_size": 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": 0.11321040242910385, "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": 0.14059995114803314, "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": 0.1292058825492859, "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": 0.12547914683818817, "qparams": { "group_size": 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": 0.10652171820402145, "qparams": { "group_size": 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": 0.0997028723359108, "qparams": { "group_size": 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": 0.07156267762184143, "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": 0.06190779432654381, "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": 0.05878153070807457, "qparams": { "group_size": 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": 0.05803323909640312, "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": 0.03588653728365898, "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": 0.030719038099050522, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030596349388360977, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02763409912586212, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027156023308634758, "qparams": { "group_size": 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": 0.01929602399468422, "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": 0.01951058954000473, "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": 0.018294380977749825, "qparams": { "group_size": 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": 0.013640553690493107, "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": 0.12226102501153946, "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": 0.12226102501153946, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1271224468946457, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11820697039365768, "qparams": { "group_size": 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": 0.11459419131278992, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10301529616117477, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05926920846104622, "qparams": { "group_size": 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": 0.055721212178468704, "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": 0.06771538406610489, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06259046494960785, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06021275371313095, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05278313159942627, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.049842339009046555, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03440871089696884, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029924584552645683, "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": 0.0284136775881052, "qparams": { "group_size": 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": 0.02805522084236145, "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": 0.01720125786960125, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014656933024525642, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014505453407764435, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013347843661904335, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01311913039535284, "qparams": { "group_size": 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": 0.008973970077931881, "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": 0.009040651842951775, "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": 0.008463682606816292, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00598269235342741, "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": 0.11820697039365768, "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": 0.11820697039365768, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.10139308869838715, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09433131664991379, "qparams": { "group_size": 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": 0.09084609895944595, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08172200620174408, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04714685305953026, "qparams": { "group_size": 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": 0.04404686018824577, "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": 0.054661825299263, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.050700731575489044, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0478864423930645, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04203423485159874, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03976782038807869, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.02766985073685646, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0241891797631979, "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": 0.022593455389142036, "qparams": { "group_size": 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": 0.02220766991376877, "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": 0.013826662674546242, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011624721810221672, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01144944317638874, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.010591626167297363, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010347001254558563, "qparams": { "group_size": 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": 0.0072064828127622604, "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": 0.007200981955975294, "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": 0.0067115649580955505, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004670246038585901, "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": 0.10139308869838715, "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": 0.10139308869838715, "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": 0.24742551147937775, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.23077552020549774, "qparams": { "group_size": 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": 0.22437304258346558, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.20206904411315918, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11613044142723083, "qparams": { "group_size": 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": 0.10948588699102402, "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": 0.13231618702411652, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12173770368099213, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11793161183595657, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.10362592339515686, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.09789355099201202, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06725841015577316, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05831821262836456, "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": 0.05562102049589157, "qparams": { "group_size": 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": 0.05497223138809204, "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": 0.03361324220895767, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0284268818795681, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02816358022391796, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.025861432775855064, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02544623613357544, "qparams": { "group_size": 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": 0.01745639555156231, "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": 0.017105603590607643, "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": 0.01650046743452549, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.010824909433722496, "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": 0.12173770368099213, "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": 0.12173770368099213, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.2438289076089859, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.22286838293075562, "qparams": { "group_size": 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": 0.21586786210536957, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.19023935496807098, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11391831934452057, "qparams": { "group_size": 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": 0.10620973259210587, "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": 0.12958751618862152, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11881613731384277, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1161244735121727, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09835269302129745, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0917196050286293, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0669369250535965, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.057822298258543015, "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": 0.0556112676858902, "qparams": { "group_size": 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": 0.05508968234062195, "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": 0.03398002311587334, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030117295682430267, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029953990131616592, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027172138914465904, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026845740154385567, "qparams": { "group_size": 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": 0.019367536529898643, "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": 0.020371414721012115, "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": 0.018702637404203415, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01581202819943428, "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": 0.11881613731384277, "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": 0.11881613731384277, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.18561625480651855, "qparams": { "group_size": 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": 0.1737290620803833, "qparams": { "group_size": 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": 0.1698177307844162, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.15359172224998474, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08745420724153519, "qparams": { "group_size": 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": 0.08312258124351501, "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": 0.09757234901189804, "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": 0.090323306620121, "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": 0.08872435241937637, "qparams": { "group_size": 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": 0.07839800417423248, "qparams": { "group_size": 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": 0.07419328391551971, "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": 0.049823399633169174, "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": 0.0433877632021904, "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": 0.04209472984075546, "qparams": { "group_size": 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": 0.04179498925805092, "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": 0.024917060509324074, "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": 0.022022737190127373, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021931208670139313, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.020233165472745895, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020048588514328003, "qparams": { "group_size": 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": 0.013277166523039341, "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": 0.013862716034054756, "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": 0.012860432267189026, "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": 0.009839319624006748, "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": 0.09757234901189804, "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": 0.09757234901189804, "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": 0.2522282898426056, "qparams": { "group_size": 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": 0.23623333871364594, "qparams": { "group_size": 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": 0.2312207967042923, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20914864540100098, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11862306296825409, "qparams": { "group_size": 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": 0.11292491853237152, "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": 0.13206139206886292, "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": 0.12219003587961197, "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": 0.12033989280462265, "qparams": { "group_size": 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": 0.10644187778234482, "qparams": { "group_size": 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": 0.10054469853639603, "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": 0.06712345778942108, "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": 0.058376457542181015, "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": 0.0567711666226387, "qparams": { "group_size": 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": 0.056395288556814194, "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": 0.03350532799959183, "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": 0.028936706483364105, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028822243213653564, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02640952728688717, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026167424395680428, "qparams": { "group_size": 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": 0.017329823225736618, "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": 0.017069999128580093, "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": 0.016773153096437454, "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": 0.010748349130153656, "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": 0.12219003587961197, "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": 0.12033989280462265, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "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": 0.26601797342300415, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24056676030158997, "qparams": { "group_size": 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": 0.23158957064151764, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20548532903194427, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12259665131568909, "qparams": { "group_size": 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": 0.1131296381354332, "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": 0.1411161720752716, "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": 0.13000456988811493, "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": 0.1260114163160324, "qparams": { "group_size": 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": 0.10642829537391663, "qparams": { "group_size": 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": 0.09944435954093933, "qparams": { "group_size": 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": 0.07191338390111923, "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": 0.062277015298604965, "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": 0.05890972539782524, "qparams": { "group_size": 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": 0.05809320509433746, "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": 0.03608977049589157, "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": 0.03071459010243416, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030572298914194107, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027531258761882782, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027011821046471596, "qparams": { "group_size": 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": 0.019360138103365898, "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": 0.019454123452305794, "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": 0.01828630082309246, "qparams": { "group_size": 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": 0.013458624482154846, "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": 0.12259665131568909, "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": 0.12259665131568909, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.12147025018930435, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11304762214422226, "qparams": { "group_size": 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": 0.10956388711929321, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09859662503004074, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0566827692091465, "qparams": { "group_size": 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": 0.05329864099621773, "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": 0.06477759033441544, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05986259877681732, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05760306864976883, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.050547484308481216, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04781355336308479, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03293316811323166, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.028637003153562546, "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": 0.02718169428408146, "qparams": { "group_size": 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": 0.026828328147530556, "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": 0.01647239364683628, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014005620032548904, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0138601865619421, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012764445506036282, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012541287578642368, "qparams": { "group_size": 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": 0.008608206175267696, "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": 0.008619971573352814, "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": 0.008125124499201775, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005673510953783989, "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": 0.12147025018930435, "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": 0.12147025018930435, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.10041514784097672, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09345366060733795, "qparams": { "group_size": 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": 0.09005089849233627, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08098620176315308, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04669766128063202, "qparams": { "group_size": 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": 0.043683867901563644, "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": 0.05404120311141014, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05016256123781204, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04745732247829437, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.041665948927402496, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.03940131142735481, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.027376238256692886, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023933295160531998, "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": 0.02239169552922249, "qparams": { "group_size": 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": 0.02201717346906662, "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": 0.013679965399205685, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011541450396180153, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011372271925210953, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.010520230047404766, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010284576565027237, "qparams": { "group_size": 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": 0.007150764111429453, "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": 0.0071680424734950066, "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": 0.006678406149148941, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0046870834194123745, "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": 0.10041514784097672, "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": 0.10041514784097672, "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": 0.25343289971351624, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.23666603863239288, "qparams": { "group_size": 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": 0.23044314980506897, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.20767255127429962, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11906027793884277, "qparams": { "group_size": 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": 0.11246541142463684, "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": 0.13514842092990875, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12427312880754471, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1208026260137558, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.10633817315101624, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10043758153915405, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06876253336668015, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.059479840099811554, "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": 0.05699997767806053, "qparams": { "group_size": 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": 0.056397195905447006, "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": 0.03435581550002098, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02914811670780182, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028906790539622307, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026542549952864647, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026157600805163383, "qparams": { "group_size": 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": 0.017841191962361336, "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": 0.017489902675151825, "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": 0.016925428062677383, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01110025029629469, "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": 0.1208026260137558, "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": 0.1208026260137558, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "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": 0.22984376549720764, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2090444564819336, "qparams": { "group_size": 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": 0.19939953088760376, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1753826141357422, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10634466260671616, "qparams": { "group_size": 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": 0.09778910130262375, "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": 0.12857788801193237, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11759187281131744, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11010808497667313, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09343177825212479, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08645152300596237, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06598831713199615, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.056896138936281204, "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": 0.05172106251120567, "qparams": { "group_size": 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": 0.05044135823845863, "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": 0.03321484103798866, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.027756422758102417, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0272822305560112, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02527189441025257, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02450130321085453, "qparams": { "group_size": 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": 0.01805364526808262, "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": 0.01892912946641445, "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": 0.016387103125452995, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01401091180741787, "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": 0.11759187281131744, "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": 0.11759187281131744, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "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": 0.18453338742256165, "qparams": { "group_size": 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": 0.17269392311573029, "qparams": { "group_size": 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": 0.16885314881801605, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.15278217196464539, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08697760105133057, "qparams": { "group_size": 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": 0.08268255740404129, "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": 0.0968988686800003, "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": 0.08975434303283691, "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": 0.08822707086801529, "qparams": { "group_size": 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": 0.07799629867076874, "qparams": { "group_size": 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": 0.07382974773645401, "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": 0.04950690641999245, "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": 0.04316578060388565, "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": 0.04191554710268974, "qparams": { "group_size": 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": 0.041614074259996414, "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": 0.024785619229078293, "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": 0.022002484649419785, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021910449489951134, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0202384814620018, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020057080313563347, "qparams": { "group_size": 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": 0.013256286270916462, "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": 0.01395753026008606, "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": 0.012848769314587116, "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": 0.010035194456577301, "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": 0.0968988686800003, "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": 0.0968988686800003, "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": 0.253836452960968, "qparams": { "group_size": 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": 0.23781654238700867, "qparams": { "group_size": 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": 0.23275381326675415, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21063312888145447, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11943909525871277, "qparams": { "group_size": 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": 0.11370556801557541, "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": 0.1329714059829712, "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": 0.12291297316551208, "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": 0.1211540699005127, "qparams": { "group_size": 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": 0.10717691481113434, "qparams": { "group_size": 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": 0.10132182389497757, "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": 0.06758097559213638, "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": 0.05873711407184601, "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": 0.057161156088113785, "qparams": { "group_size": 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": 0.05679541081190109, "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": 0.033749110996723175, "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": 0.02915245108306408, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029042545706033707, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02661476843059063, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026377275586128235, "qparams": { "group_size": 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": 0.01750919595360756, "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": 0.017211301252245903, "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": 0.016965465620160103, "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": 0.010870185680687428, "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": 0.1211540699005127, "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": 0.1211540699005127, "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": 0.2667391002178192, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24044787883758545, "qparams": { "group_size": 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": 0.23105405271053314, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2048095166683197, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12279347330331802, "qparams": { "group_size": 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": 0.11293356120586395, "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": 0.14196358621120453, "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": 0.13073617219924927, "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": 0.12643548846244812, "qparams": { "group_size": 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": 0.1063498705625534, "qparams": { "group_size": 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": 0.0994938313961029, "qparams": { "group_size": 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": 0.07245147228240967, "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": 0.06267601251602173, "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": 0.05906255543231964, "qparams": { "group_size": 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": 0.05820336192846298, "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": 0.03639282286167145, "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": 0.030880585312843323, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03071717731654644, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027651144191622734, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02708272449672222, "qparams": { "group_size": 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": 0.01963101513683796, "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": 0.01970076560974121, "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": 0.018477261066436768, "qparams": { "group_size": 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": 0.013710192404687405, "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": 0.12279347330331802, "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": 0.12279347330331802, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.12858040630817413, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11980870366096497, "qparams": { "group_size": 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": 0.11628955602645874, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10469979047775269, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.060059092938899994, "qparams": { "group_size": 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": 0.05660121887922287, "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": 0.06849071383476257, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06326691806316376, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.061022933572530746, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.053635429590940475, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.050753794610500336, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.034851234406232834, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.030280696228146553, "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": 0.028844205662608147, "qparams": { "group_size": 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": 0.028502412140369415, "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": 0.017443977296352386, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01494024321436882, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014799654483795166, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013648765161633492, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013433011248707771, "qparams": { "group_size": 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": 0.009158791042864323, "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": 0.009283853694796562, "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": 0.00868210382759571, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006268636789172888, "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": 0.11980870366096497, "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": 0.11980870366096497, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.10424517095088959, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09706360846757889, "qparams": { "group_size": 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": 0.09361767768859863, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08429694920778275, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04851805791258812, "qparams": { "group_size": 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": 0.04545975476503372, "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": 0.055993419140577316, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05196632072329521, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04929398372769356, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04330269992351532, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0409436896443367, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.028351841494441032, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02479959838092327, "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": 0.023268872871994972, "qparams": { "group_size": 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": 0.022895419970154762, "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": 0.014177830889821053, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011992847546935081, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011824599467217922, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01093879621475935, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01070209126919508, "qparams": { "group_size": 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": 0.007419670931994915, "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": 0.007434719242155552, "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": 0.006946190260350704, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004862431436777115, "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": 0.10424517095088959, "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": 0.10424517095088959, "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": 0.26215317845344543, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.24514667689800262, "qparams": { "group_size": 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": 0.23885193467140198, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.2156379669904709, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.123322494328022, "qparams": { "group_size": 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": 0.11668704450130463, "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": 0.13946683704853058, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12839026749134064, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12513041496276855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11027731001377106, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10433124005794525, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0709332674741745, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.061423808336257935, "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": 0.05901080742478371, "qparams": { "group_size": 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": 0.05844935402274132, "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": 0.03540796414017677, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030117427930235863, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029896443709731102, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027438780292868614, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027069751173257828, "qparams": { "group_size": 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": 0.018313737586140633, "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": 0.01795096881687641, "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": 0.017428353428840637, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011270508170127869, "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": 0.11027731001377106, "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": 0.11027731001377106, "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": 0.23531854152679443, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21432094275951385, "qparams": { "group_size": 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": 0.20683050155639648, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1800634264945984, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10873261839151382, "qparams": { "group_size": 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": 0.100632444024086, "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": 0.12496696412563324, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11446017026901245, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11147312074899673, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09305400401353836, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08582084625959396, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06422547996044159, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05571891739964485, "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": 0.053105250000953674, "qparams": { "group_size": 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": 0.052490439265966415, "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": 0.03256966546177864, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.028951844200491905, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028788037598133087, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026035627350211143, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025668904185295105, "qparams": { "group_size": 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": 0.018343882635235786, "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": 0.019871246069669724, "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": 0.017529543489217758, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015554552897810936, "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": 0.11446017026901245, "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": 0.11446017026901245, "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": 0.1802869439125061, "qparams": { "group_size": 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": 0.16876207292079926, "qparams": { "group_size": 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": 0.1649874448776245, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1493540108203888, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.0848914086818695, "qparams": { "group_size": 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": 0.08072873204946518, "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": 0.0947021096944809, "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": 0.08759969472885132, "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": 0.08611119538545609, "qparams": { "group_size": 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": 0.07614454627037048, "qparams": { "group_size": 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": 0.07212452590465546, "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": 0.048291437327861786, "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": 0.04208424687385559, "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": 0.040863655507564545, "qparams": { "group_size": 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": 0.04057150334119797, "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": 0.024185528978705406, "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": 0.021376246586441994, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021286746487021446, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019648805260658264, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.019473770633339882, "qparams": { "group_size": 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": 0.012896678410470486, "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": 0.01344996877014637, "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": 0.012499861419200897, "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": 0.009545179083943367, "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": 0.0947021096944809, "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": 0.0947021096944809, "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": 0.2496495246887207, "qparams": { "group_size": 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": 0.23400886356830597, "qparams": { "group_size": 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": 0.22906644642353058, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2074604332447052, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11744576692581177, "qparams": { "group_size": 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": 0.11182106286287308, "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": 0.1306666135787964, "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": 0.12080222368240356, "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": 0.11907940357923508, "qparams": { "group_size": 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": 0.10542918741703033, "qparams": { "group_size": 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": 0.09977283328771591, "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": 0.06645063310861588, "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": 0.057722438126802444, "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": 0.056178558617830276, "qparams": { "group_size": 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": 0.055822573602199554, "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": 0.033215444535017014, "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": 0.02865067683160305, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028542660176753998, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02617303654551506, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0259388517588377, "qparams": { "group_size": 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": 0.017288079485297203, "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": 0.016906844452023506, "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": 0.01677081547677517, "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": 0.010668759234249592, "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": 0.12080222368240356, "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": 0.12080222368240356, "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": 0.2625572979450226, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23666618764400482, "qparams": { "group_size": 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": 0.22717507183551788, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20151694118976593, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12072044610977173, "qparams": { "group_size": 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": 0.11099627614021301, "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": 0.139851912856102, "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": 0.12880559265613556, "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": 0.12427029013633728, "qparams": { "group_size": 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": 0.10460999608039856, "qparams": { "group_size": 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": 0.09796058386564255, "qparams": { "group_size": 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": 0.07133279740810394, "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": 0.0616663359105587, "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": 0.058016758412122726, "qparams": { "group_size": 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": 0.05714070796966553, "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": 0.03583309426903725, "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": 0.030261533334851265, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030090907588601112, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027082964777946472, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02652382291853428, "qparams": { "group_size": 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": 0.019331028684973717, "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": 0.01922607421875, "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": 0.018175490200519562, "qparams": { "group_size": 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": 0.01328637357801199, "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": 0.12072044610977173, "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": 0.12072044610977173, "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": 0.126617431640625, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11808580160140991, "qparams": { "group_size": 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": 0.11461130529642105, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10327155888080597, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05911043658852577, "qparams": { "group_size": 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": 0.05572567507624626, "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": 0.0675593689084053, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.062318045645952225, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06000468134880066, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05284537002444267, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.050129346549510956, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.034349650144577026, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029788365587592125, "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": 0.028333967551589012, "qparams": { "group_size": 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": 0.027993181720376015, "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": 0.017170583829283714, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014599348418414593, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014456669799983501, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013330425135791302, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013113502413034439, "qparams": { "group_size": 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": 0.008969169110059738, "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": 0.008961435407400131, "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": 0.008484621532261372, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005893946159631014, "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": 0.11808580160140991, "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": 0.11808580160140991, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1034516915678978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09638868272304535, "qparams": { "group_size": 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": 0.09296466410160065, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08371265232563019, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04816528409719467, "qparams": { "group_size": 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": 0.04509998857975006, "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": 0.05574033409357071, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.051675934344530106, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04891492798924446, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04308054968714714, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04078981652855873, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.028248893097043037, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.024684855714440346, "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": 0.023105382919311523, "qparams": { "group_size": 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": 0.022720931097865105, "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": 0.014120279811322689, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011925149708986282, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011753224767744541, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.010897969827055931, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010654940269887447, "qparams": { "group_size": 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": 0.007395926397293806, "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": 0.0074291895143687725, "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": 0.006913834251463413, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004897044971585274, "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": 0.1034516915678978, "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": 0.1034516915678978, "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": 0.2609327435493469, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.24420222640037537, "qparams": { "group_size": 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": 0.2381971925497055, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.21520192921161652, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12264131754636765, "qparams": { "group_size": 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": 0.11625175178050995, "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": 0.13847120106220245, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12747643887996674, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12442515790462494, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.10986465215682983, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10401187837123871, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0704052746295929, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06091984361410141, "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": 0.05869295820593834, "qparams": { "group_size": 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": 0.05817432701587677, "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": 0.03515283018350601, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.029960885643959045, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029761413112282753, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02732636220753193, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02698599174618721, "qparams": { "group_size": 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": 0.018169958144426346, "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": 0.01783103495836258, "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": 0.017329726368188858, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011247999034821987, "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": 0.12264131754636765, "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": 0.12264131754636765, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.24076157808303833, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21446183323860168, "qparams": { "group_size": 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": 0.20599773526191711, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.17318607866764069, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1123068630695343, "qparams": { "group_size": 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": 0.10276881605386734, "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": 0.12881791591644287, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11818110197782516, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11504270136356354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0928594172000885, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08312266319990158, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06611748784780502, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.057349514216184616, "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": 0.0546921044588089, "qparams": { "group_size": 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": 0.05405734106898308, "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": 0.033276915550231934, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0294693261384964, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029272837564349174, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02579457312822342, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02538076415657997, "qparams": { "group_size": 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": 0.01837010309100151, "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": 0.019828449934720993, "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": 0.017576241865754128, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015140094794332981, "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": 0.11818110197782516, "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": 0.11818110197782516, "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": 0.18918192386627197, "qparams": { "group_size": 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": 0.17712725698947906, "qparams": { "group_size": 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": 0.17322686314582825, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.15678109228610992, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08909492194652557, "qparams": { "group_size": 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": 0.08473172783851624, "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": 0.09924500435590744, "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": 0.09182308614253998, "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": 0.09037856012582779, "qparams": { "group_size": 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": 0.07991740107536316, "qparams": { "group_size": 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": 0.07563956081867218, "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": 0.05059744045138359, "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": 0.044068146497011185, "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": 0.042825717478990555, "qparams": { "group_size": 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": 0.04253431409597397, "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": 0.025361480191349983, "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": 0.022294484078884125, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022204143926501274, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.020465299487113953, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020280340686440468, "qparams": { "group_size": 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": 0.013518957421183586, "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": 0.013863349333405495, "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": 0.01311973575502634, "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": 0.009659492410719395, "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": 0.09924500435590744, "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": 0.09924500435590744, "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": 0.2582310736179352, "qparams": { "group_size": 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": 0.2420331984758377, "qparams": { "group_size": 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": 0.23694999516010284, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21459989249706268, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12143594026565552, "qparams": { "group_size": 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": 0.11570427566766739, "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": 0.13521793484687805, "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": 0.12491617351770401, "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": 0.12317360937595367, "qparams": { "group_size": 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": 0.10906156897544861, "qparams": { "group_size": 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": 0.10319149494171143, "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": 0.06869882345199585, "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": 0.05968761444091797, "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": 0.05813154950737953, "qparams": { "group_size": 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": 0.057757630944252014, "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": 0.034388378262519836, "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": 0.02966381423175335, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029558705165982246, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027118539437651634, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026878930628299713, "qparams": { "group_size": 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": 0.01802724599838257, "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": 0.01754479482769966, "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": 0.01751135289669037, "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": 0.01113482192158699, "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": 0.12143594026565552, "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": 0.12143594026565552, "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": 0.2665615975856781, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.239729642868042, "qparams": { "group_size": 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": 0.2297932356595993, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20363812148571014, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12250315397977829, "qparams": { "group_size": 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": 0.11231418699026108, "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": 0.14209970831871033, "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": 0.13105551898479462, "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": 0.1262161135673523, "qparams": { "group_size": 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": 0.1059441864490509, "qparams": { "group_size": 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": 0.09921536594629288, "qparams": { "group_size": 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": 0.0724465474486351, "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": 0.06289467960596085, "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": 0.05901334062218666, "qparams": { "group_size": 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": 0.05805385485291481, "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": 0.03637049347162247, "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": 0.030981019139289856, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030783556401729584, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027747854590415955, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02714364416897297, "qparams": { "group_size": 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": 0.019387558102607727, "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": 0.01998104527592659, "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": 0.018080363050103188, "qparams": { "group_size": 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": 0.01408088207244873, "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": 0.12250315397977829, "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": 0.12250315397977829, "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": 0.12923195958137512, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12070260941982269, "qparams": { "group_size": 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": 0.11722727119922638, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10575862228870392, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06045285612344742, "qparams": { "group_size": 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": 0.057047586888074875, "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": 0.06897246837615967, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06365431845188141, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.061351947486400604, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.054112888872623444, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.051361773163080215, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03506876900792122, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03042316623032093, "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": 0.028963379561901093, "qparams": { "group_size": 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": 0.028613310307264328, "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": 0.017544738948345184, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01490702386945486, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01476428285241127, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013631158508360386, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013410333544015884, "qparams": { "group_size": 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": 0.009206344373524189, "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": 0.009134397841989994, "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": 0.008721565827727318, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005994528532028198, "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": 0.12070260941982269, "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": 0.12070260941982269, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.10855467617511749, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1012713834643364, "qparams": { "group_size": 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": 0.09788242727518082, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08825565129518509, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05061212182044983, "qparams": { "group_size": 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": 0.04753852263092995, "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": 0.058318283408880234, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05399589613080025, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05139674246311188, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04530792683362961, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04293587803840637, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029534833505749702, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025774849578738213, "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": 0.024261629208922386, "qparams": { "group_size": 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": 0.023901235312223434, "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": 0.014762947335839272, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012486926279962063, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012322750873863697, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011414160020649433, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011181030422449112, "qparams": { "group_size": 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": 0.007711655925959349, "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": 0.007700592279434204, "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": 0.007239505182951689, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005018861033022404, "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": 0.10855467617511749, "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": 0.10855467617511749, "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": 0.2663295269012451, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.24927039444446564, "qparams": { "group_size": 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": 0.2434585839509964, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.2200443297624588, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12519386410713196, "qparams": { "group_size": 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": 0.11886214464902878, "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": 0.14085732400417328, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12972085177898407, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1270069032907486, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11223010718822479, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10619816184043884, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07158602774143219, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06197816878557205, "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": 0.059878453612327576, "qparams": { "group_size": 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": 0.05938588082790375, "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": 0.035739827901124954, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030446315184235573, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03026745282113552, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027765145525336266, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02744472771883011, "qparams": { "group_size": 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": 0.01838456094264984, "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": 0.017933746799826622, "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": 0.01758296974003315, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011074072681367397, "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": 0.11223010718822479, "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": 0.11223010718822479, "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": 0.2509366571903229, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.22047863900661469, "qparams": { "group_size": 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": 0.2107909917831421, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.17785391211509705, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11711350083351135, "qparams": { "group_size": 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": 0.10598620772361755, "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": 0.13378192484378815, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12303659319877625, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12010233104228973, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09470837563276291, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08658821880817413, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0690813884139061, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05995584651827812, "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": 0.057320013642311096, "qparams": { "group_size": 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": 0.056697066873311996, "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": 0.035069819539785385, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.031310178339481354, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03114648163318634, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027151966467499733, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026746001094579697, "qparams": { "group_size": 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": 0.02004455216228962, "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": 0.02153380587697029, "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": 0.019292093813419342, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.016942407935857773, "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": 0.12010233104228973, "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": 0.12010233104228973, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "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": 0.1962576061487198, "qparams": { "group_size": 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": 0.18387433886528015, "qparams": { "group_size": 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": 0.17992419004440308, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.16277074813842773, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09240971505641937, "qparams": { "group_size": 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": 0.08797570317983627, "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": 0.10274078696966171, "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": 0.09518048167228699, "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": 0.09373950213193893, "qparams": { "group_size": 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": 0.08293530344963074, "qparams": { "group_size": 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": 0.07841269671916962, "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": 0.05238446593284607, "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": 0.045646730810403824, "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": 0.04439862072467804, "qparams": { "group_size": 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": 0.04410424456000328, "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": 0.02618648111820221, "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": 0.023051710799336433, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02296612225472927, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021162211894989014, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020976023748517036, "qparams": { "group_size": 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": 0.0138392960652709, "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": 0.014255602844059467, "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": 0.013433066196739674, "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": 0.009850304573774338, "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": 0.10274078696966171, "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": 0.10274078696966171, "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": 0.2641589045524597, "qparams": { "group_size": 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": 0.2477215677499771, "qparams": { "group_size": 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": 0.2425985187292099, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21971237659454346, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1243351623415947, "qparams": { "group_size": 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": 0.11848670989274979, "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": 0.13790498673915863, "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": 0.1277412325143814, "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": 0.12608017027378082, "qparams": { "group_size": 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": 0.11163796484470367, "qparams": { "group_size": 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": 0.1054929792881012, "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": 0.07009825855493546, "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": 0.06102416664361954, "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": 0.059462420642375946, "qparams": { "group_size": 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": 0.05909125134348869, "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": 0.03498070687055588, "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": 0.030296653509140015, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03019075095653534, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027681540697813034, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027444588020443916, "qparams": { "group_size": 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": 0.018063480034470558, "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": 0.01783723756670952, "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": 0.017536073923110962, "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": 0.011204841546714306, "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": 0.11163796484470367, "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": 0.11163796484470367, "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": 0.26915982365608215, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.2415245771408081, "qparams": { "group_size": 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": 0.23120807111263275, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20498859882354736, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12352439761161804, "qparams": { "group_size": 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": 0.11304384469985962, "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": 0.14383098483085632, "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": 0.13260091841220856, "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": 0.12745504081249237, "qparams": { "group_size": 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": 0.10673948377370834, "qparams": { "group_size": 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": 0.10009558498859406, "qparams": { "group_size": 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": 0.07320617139339447, "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": 0.06349984556436539, "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": 0.05939112603664398, "qparams": { "group_size": 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": 0.05839390680193901, "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": 0.036639921367168427, "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": 0.030960915610194206, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030755354091525078, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027644524350762367, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02699786052107811, "qparams": { "group_size": 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": 0.01945047825574875, "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": 0.019718758761882782, "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": 0.018107784911990166, "qparams": { "group_size": 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": 0.013532610610127449, "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": 0.10673948377370834, "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": 0.10673948377370834, "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": 0.12731583416461945, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11885262280702591, "qparams": { "group_size": 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": 0.11523592472076416, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10396717488765717, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.059512898325920105, "qparams": { "group_size": 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": 0.056083086878061295, "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": 0.06813112646341324, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06286358833312988, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.060418665409088135, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.053290680050849915, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.050536610186100006, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.034658681601285934, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.030072614550590515, "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": 0.02855944074690342, "qparams": { "group_size": 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": 0.02819846197962761, "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": 0.017343688756227493, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014768757857382298, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01461589802056551, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013515550643205643, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01328447088599205, "qparams": { "group_size": 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": 0.009116084314882755, "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": 0.009158501401543617, "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": 0.008626015856862068, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006120815873146057, "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": 0.11885262280702591, "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": 0.11885262280702591, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.10765539109706879, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10037810355424881, "qparams": { "group_size": 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": 0.0968344658613205, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08730096369981766, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05018628388643265, "qparams": { "group_size": 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": 0.047023575752973557, "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": 0.05806925892829895, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05380956456065178, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05098017305135727, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04489031434059143, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.042576637119054794, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029417559504508972, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025693487375974655, "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": 0.024067964404821396, "qparams": { "group_size": 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": 0.023672252893447876, "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": 0.01470982190221548, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012400050647556782, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012220583856105804, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011329371482133865, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01108148880302906, "qparams": { "group_size": 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": 0.007696222979575396, "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": 0.00769097451120615, "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": 0.00719962315633893, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005028344690799713, "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": 0.10765539109706879, "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": 0.10765539109706879, "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": 0.26578834652900696, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.24868211150169373, "qparams": { "group_size": 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": 0.2427017241716385, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.2194427251815796, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12495752424001694, "qparams": { "group_size": 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": 0.11858469992876053, "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": 0.14087896049022675, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12966564297676086, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12679290771484375, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11198519170284271, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10600271821022034, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07161872088909149, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.061935968697071075, "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": 0.05976902320981026, "qparams": { "group_size": 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": 0.05926133319735527, "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": 0.035717789083719254, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030433855950832367, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030252255499362946, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02775198593735695, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027423754334449768, "qparams": { "group_size": 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": 0.01841169409453869, "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": 0.01799697056412697, "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": 0.017580419778823853, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01122397929430008, "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": 0.11198519170284271, "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": 0.11198519170284271, "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": 0.23752461373806, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.21118640899658203, "qparams": { "group_size": 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": 0.20232555270195007, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.17393344640731812, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11022025346755981, "qparams": { "group_size": 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": 0.10014311224222183, "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": 0.12783312797546387, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11691368371248245, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1135031208395958, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09228038787841797, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08421143144369125, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06602564454078674, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05682054162025452, "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": 0.053772684186697006, "qparams": { "group_size": 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": 0.05303728207945824, "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": 0.03349180147051811, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.029252909123897552, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02906680479645729, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02592560090124607, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025471501052379608, "qparams": { "group_size": 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": 0.01908409595489502, "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": 0.020054113119840622, "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": 0.018191026523709297, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015599509701132774, "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": 0.11691368371248245, "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": 0.11691368371248245, "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": 0.2013908326625824, "qparams": { "group_size": 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": 0.18875671923160553, "qparams": { "group_size": 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": 0.18474140763282776, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1672457605600357, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09485503286123276, "qparams": { "group_size": 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": 0.09029311686754227, "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": 0.10560929775238037, "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": 0.09764143079519272, "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": 0.09622178971767426, "qparams": { "group_size": 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": 0.08516237139701843, "qparams": { "group_size": 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": 0.08066476136445999, "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": 0.05379775911569595, "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": 0.046852827072143555, "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": 0.04559123516082764, "qparams": { "group_size": 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": 0.04530133679509163, "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": 0.026963740587234497, "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": 0.02373649924993515, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023651542142033577, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021801091730594635, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021619100123643875, "qparams": { "group_size": 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": 0.01438211090862751, "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": 0.014753419905900955, "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": 0.013974819332361221, "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": 0.010288198478519917, "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": 0.10560929775238037, "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": 0.10560929775238037, "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": 0.26517021656036377, "qparams": { "group_size": 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": 0.24864763021469116, "qparams": { "group_size": 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": 0.24356748163700104, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.22055643796920776, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.124766506254673, "qparams": { "group_size": 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": 0.1189192458987236, "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": 0.13851045072078705, "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": 0.1281842291355133, "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": 0.12651751935482025, "qparams": { "group_size": 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": 0.11206799000501633, "qparams": { "group_size": 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": 0.10601592063903809, "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": 0.07042022049427032, "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": 0.06127274036407471, "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": 0.05970780551433563, "qparams": { "group_size": 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": 0.05934109911322594, "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": 0.03521955385804176, "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": 0.03045344352722168, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030348224565386772, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027834920212626457, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027599604800343513, "qparams": { "group_size": 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": 0.018384814262390137, "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": 0.017968036234378815, "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": 0.017870759591460228, "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": 0.011363453231751919, "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": 0.11206799000501633, "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": 0.11206799000501633, "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": 0.2666179835796356, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.2379421889781952, "qparams": { "group_size": 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": 0.22699274122714996, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2008979469537735, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12215259671211243, "qparams": { "group_size": 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": 0.11115457862615585, "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": 0.1429090052843094, "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": 0.13184882700443268, "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": 0.12627336382865906, "qparams": { "group_size": 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": 0.1049557477235794, "qparams": { "group_size": 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": 0.09830329567193985, "qparams": { "group_size": 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": 0.0727381706237793, "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": 0.06304047256708145, "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": 0.05867889150977135, "qparams": { "group_size": 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": 0.05761533975601196, "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": 0.03635025769472122, "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": 0.0305266585201025, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03029850870370865, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027128644287586212, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026435086503624916, "qparams": { "group_size": 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": 0.0192565880715847, "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": 0.01940624974668026, "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": 0.017847269773483276, "qparams": { "group_size": 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": 0.013169311918318272, "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": 0.12215259671211243, "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": 0.12215259671211243, "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": 0.12384527921676636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11555343121290207, "qparams": { "group_size": 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": 0.1119823232293129, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10110142827033997, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05777374655008316, "qparams": { "group_size": 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": 0.05442220717668533, "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": 0.06618589907884598, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06116671860218048, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.058676209300756454, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05173293873667717, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04908499866724014, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.033578503876924515, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0292021743953228, "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": 0.027685198932886124, "qparams": { "group_size": 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": 0.027317941188812256, "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": 0.016796160489320755, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014233971014618874, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01408008486032486, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013007848523557186, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012774471193552017, "qparams": { "group_size": 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": 0.008785983547568321, "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": 0.008726613596081734, "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": 0.008295557461678982, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005684390664100647, "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": 0.11555343121290207, "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": 0.11555343121290207, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.1050264835357666, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09785803407430649, "qparams": { "group_size": 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": 0.09439119696617126, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08511777967214584, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04887891933321953, "qparams": { "group_size": 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": 0.04579828307032585, "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": 0.056650418788194656, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05247743800282478, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.04967128485441208, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04375792667269707, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.041500892490148544, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.028682604432106018, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02505986951291561, "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": 0.023440947756171227, "qparams": { "group_size": 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": 0.023049892857670784, "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": 0.014348853379487991, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012079870328307152, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011905715800821781, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011042746715247631, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010796560905873775, "qparams": { "group_size": 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": 0.00750606507062912, "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": 0.007499292958527803, "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": 0.007014047354459763, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00490416819229722, "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": 0.1050264835357666, "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": 0.1050264835357666, "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": 0.25701767206192017, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.24029920995235443, "qparams": { "group_size": 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": 0.23428575694561005, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.21168792247772217, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12048937380313873, "qparams": { "group_size": 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": 0.11417432129383087, "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": 0.136166051030159, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1253654658794403, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12227707356214523, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.1079535186290741, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10220303386449814, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.069163978099823, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05986862629652023, "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": 0.057613857090473175, "qparams": { "group_size": 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": 0.057084400206804276, "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": 0.03450557589530945, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02934800088405609, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029151676222682, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026767395436763763, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02642006427049637, "qparams": { "group_size": 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": 0.017806196585297585, "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": 0.01740085706114769, "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": 0.01697397790849209, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.010846513323485851, "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": 0.12227707356214523, "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": 0.12048937380313873, "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": 0.2400369793176651, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20903225243091583, "qparams": { "group_size": 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": 0.19781120121479034, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1658947616815567, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.11049727350473404, "qparams": { "group_size": 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": 0.09876001626253128, "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": 0.13086360692977905, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1199442446231842, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11498437076807022, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.09020785242319107, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08120542764663696, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0673690140247345, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.058229342103004456, "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": 0.053934138268232346, "qparams": { "group_size": 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": 0.05288674309849739, "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": 0.0341973751783371, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.029353071004152298, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02903877943754196, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.025602789595723152, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02493460662662983, "qparams": { "group_size": 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": 0.019391700625419617, "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": 0.020337937399744987, "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": 0.01810392737388611, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015659112483263016, "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": 0.1199442446231842, "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": 0.1199442446231842, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20170991122722626, "qparams": { "group_size": 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": 0.18913617730140686, "qparams": { "group_size": 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": 0.1850825995206833, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1675451546907425, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09490326792001724, "qparams": { "group_size": 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": 0.0904354676604271, "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": 0.10556401312351227, "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": 0.09768494963645935, "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": 0.09625310450792313, "qparams": { "group_size": 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": 0.08525189012289047, "qparams": { "group_size": 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": 0.08067592233419418, "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": 0.05381220951676369, "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": 0.04683643952012062, "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": 0.045591678470373154, "qparams": { "group_size": 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": 0.04529101401567459, "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": 0.02695619687438011, "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": 0.023670116439461708, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023581799119710922, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021736178547143936, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021552927792072296, "qparams": { "group_size": 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": 0.014330501668155193, "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": 0.014619054272770882, "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": 0.013922255486249924, "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": 0.010092519223690033, "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": 0.10556401312351227, "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": 0.10556401312351227, "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": 0.26150748133659363, "qparams": { "group_size": 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": 0.24528272449970245, "qparams": { "group_size": 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": 0.240282341837883, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21773011982440948, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12292365729808807, "qparams": { "group_size": 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": 0.11722105741500854, "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": 0.13656793534755707, "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": 0.12631702423095703, "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": 0.12466173619031906, "qparams": { "group_size": 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": 0.11051769554615021, "qparams": { "group_size": 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": 0.10465029627084732, "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": 0.06938157975673676, "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": 0.06036486476659775, "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": 0.05883447825908661, "qparams": { "group_size": 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": 0.05846590921282768, "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": 0.03466685116291046, "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": 0.03003479354083538, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02993306703865528, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02747959829866886, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027244625613093376, "qparams": { "group_size": 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": 0.018029367551207542, "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": 0.01777650974690914, "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": 0.017510849982500076, "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": 0.011310242116451263, "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": 0.11051769554615021, "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": 0.11051769554615021, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.2669225037097931, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.2380203902721405, "qparams": { "group_size": 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": 0.22694769501686096, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2012113630771637, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12241265177726746, "qparams": { "group_size": 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": 0.11130426824092865, "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": 0.14396114647388458, "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": 0.13221755623817444, "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": 0.12658996880054474, "qparams": { "group_size": 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": 0.10518395155668259, "qparams": { "group_size": 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": 0.0987996757030487, "qparams": { "group_size": 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": 0.07329852879047394, "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": 0.0634082555770874, "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": 0.05899461358785629, "qparams": { "group_size": 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": 0.05791884660720825, "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": 0.03676744922995567, "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": 0.031018907204270363, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03078744374215603, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027668187394738197, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026977885514497757, "qparams": { "group_size": 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": 0.01979641430079937, "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": 0.0201475340873003, "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": 0.01836068369448185, "qparams": { "group_size": 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": 0.014206062071025372, "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": 0.12241265177726746, "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": 0.12241265177726746, "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": 0.1272784322500229, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1187194436788559, "qparams": { "group_size": 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": 0.11515285819768906, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10387660562992096, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.059383101761341095, "qparams": { "group_size": 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": 0.055958595126867294, "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": 0.06802336871623993, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0627608597278595, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06030024588108063, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0531817264854908, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05047200992703438, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.034569043666124344, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029999084770679474, "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": 0.0284714475274086, "qparams": { "group_size": 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": 0.028106005862355232, "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": 0.0173085555434227, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014674236066639423, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014519806951284409, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013421619310975075, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013189129531383514, "qparams": { "group_size": 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": 0.009094358421862125, "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": 0.009033232927322388, "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": 0.008603931404650211, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005944677162915468, "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": 0.1187194436788559, "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": 0.1187194436788559, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.10630784928798676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09908957034349442, "qparams": { "group_size": 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": 0.09542272984981537, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08605749905109406, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04944257810711861, "qparams": { "group_size": 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": 0.04623638466000557, "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": 0.05754288658499718, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.053315192461013794, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05020635202527046, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04426256939768791, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.042019911110401154, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0291452519595623, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025442542508244514, "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": 0.02371837943792343, "qparams": { "group_size": 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": 0.023296210914850235, "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": 0.01459728367626667, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012267853133380413, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012076104059815407, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011228658258914948, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010958355851471424, "qparams": { "group_size": 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": 0.0076974136754870415, "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": 0.007691442035138607, "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": 0.007173833437263966, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0050917924381792545, "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": 0.10630784928798676, "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": 0.10630784928798676, "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": 0.2680431604385376, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.250703364610672, "qparams": { "group_size": 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": 0.24482934176921844, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.2214345782995224, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1257546842098236, "qparams": { "group_size": 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": 0.11936776340007782, "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": 0.1415008008480072, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1302548050880432, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12756839394569397, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11269256472587585, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10669612884521484, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0719306692481041, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.062200047075748444, "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": 0.06012190505862236, "qparams": { "group_size": 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": 0.05962488427758217, "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": 0.03586765378713608, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030575793236494064, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030400820076465607, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027878325432538986, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02756180800497532, "qparams": { "group_size": 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": 0.018483977764844894, "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": 0.018016304820775986, "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": 0.017685946077108383, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011154142208397388, "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": 0.11269256472587585, "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": 0.11269256472587585, "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": 0.2188473492860794, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19076034426689148, "qparams": { "group_size": 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": 0.18139970302581787, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15106217563152313, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10134799778461456, "qparams": { "group_size": 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": 0.09119155257940292, "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": 0.11724825203418732, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10739687830209732, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1039285734295845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08066460490226746, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07426836341619492, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0602308064699173, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05236732214689255, "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": 0.04962730407714844, "qparams": { "group_size": 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": 0.04897131398320198, "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": 0.03057115152478218, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.027240129187703133, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027042608708143234, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02347620762884617, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02305673249065876, "qparams": { "group_size": 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": 0.017421426251530647, "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": 0.018960438668727875, "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": 0.01660740189254284, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01502781081944704, "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": 0.11724825203418732, "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": 0.11724825203418732, "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": 0.19411347806453705, "qparams": { "group_size": 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": 0.18201379477977753, "qparams": { "group_size": 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": 0.17818856239318848, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.16141684353351593, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09125839173793793, "qparams": { "group_size": 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": 0.0869433581829071, "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": 0.10133904963731766, "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": 0.09385290741920471, "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": 0.09250951558351517, "qparams": { "group_size": 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": 0.08200562000274658, "qparams": { "group_size": 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": 0.07765647768974304, "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": 0.051578979939222336, "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": 0.04493667185306549, "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": 0.04376396909356117, "qparams": { "group_size": 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": 0.04348299652338028, "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": 0.025788260623812675, "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": 0.022630605846643448, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02255082130432129, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.020779479295015335, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020601743832230568, "qparams": { "group_size": 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": 0.013584377244114876, "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": 0.013852955773472786, "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": 0.01319856196641922, "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": 0.00941050797700882, "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": 0.10133904963731766, "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": 0.10133904963731766, "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": 0.25295010209083557, "qparams": { "group_size": 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": 0.23746179044246674, "qparams": { "group_size": 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": 0.23263591527938843, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21088312566280365, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11888741701841354, "qparams": { "group_size": 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": 0.11340823024511337, "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": 0.1317581832408905, "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": 0.12208741903305054, "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": 0.1205228865146637, "qparams": { "group_size": 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": 0.10692544281482697, "qparams": { "group_size": 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": 0.10118981450796127, "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": 0.06697297841310501, "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": 0.058277428150177, "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": 0.05682002380490303, "qparams": { "group_size": 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": 0.05647531524300575, "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": 0.03340395539999008, "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": 0.028901370242238045, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028806645423173904, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02644190564751625, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02622044086456299, "qparams": { "group_size": 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": 0.017243171110749245, "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": 0.016941219568252563, "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": 0.01675345189869404, "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": 0.010562011040747166, "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": 0.12208741903305054, "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": 0.1205228865146637, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "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": 0.2588557302951813, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.230800062417984, "qparams": { "group_size": 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": 0.22037261724472046, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19597896933555603, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11834602802991867, "qparams": { "group_size": 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": 0.1077749952673912, "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": 0.13836677372455597, "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": 0.1275109499692917, "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": 0.12241800874471664, "qparams": { "group_size": 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": 0.10202977061271667, "qparams": { "group_size": 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": 0.09595605731010437, "qparams": { "group_size": 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": 0.07044496387243271, "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": 0.06103114038705826, "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": 0.05691375583410263, "qparams": { "group_size": 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": 0.055917490273714066, "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": 0.035302430391311646, "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": 0.029719529673457146, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029521776363253593, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02650575339794159, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025868389755487442, "qparams": { "group_size": 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": 0.018906505778431892, "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": 0.019019128754734993, "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": 0.017581989988684654, "qparams": { "group_size": 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": 0.01312279887497425, "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": 0.12241800874471664, "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": 0.12241800874471664, "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": 0.1324862241744995, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12389805167913437, "qparams": { "group_size": 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": 0.12049219757318497, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1089903935790062, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06190766394138336, "qparams": { "group_size": 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": 0.058577150106430054, "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": 0.07017897069454193, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06501361727714539, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06284444034099579, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05558949336409569, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05269559472799301, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03562397509813309, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03102409839630127, "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": 0.029641713947057724, "qparams": { "group_size": 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": 0.029312647879123688, "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": 0.01781056821346283, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.015181679278612137, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015045816078782082, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013890055939555168, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01367935724556446, "qparams": { "group_size": 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": 0.009281383827328682, "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": 0.009169998578727245, "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": 0.008833414874970913, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0058759362436831, "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": 0.12049219757318497, "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": 0.12049219757318497, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.1163109764456749, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10871053487062454, "qparams": { "group_size": 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": 0.10548018664121628, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0953739732503891, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.054289560765028, "qparams": { "group_size": 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": 0.05122436210513115, "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": 0.062063757330179214, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05738191679120064, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05510526895523071, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04874458163976669, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.046270694583654404, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03146987035870552, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02737986482679844, "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": 0.025990379974246025, "qparams": { "group_size": 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": 0.02565109170973301, "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": 0.01572338677942753, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013311750255525112, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01316784881055355, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012182088568806648, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01197030395269394, "qparams": { "group_size": 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": 0.008172713220119476, "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": 0.00807726290076971, "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": 0.007725004572421312, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005163141526281834, "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": 0.1163109764456749, "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": 0.1163109764456749, "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": 0.2700871527194977, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2532511055469513, "qparams": { "group_size": 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": 0.24770358204841614, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.22432193160057068, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12690594792366028, "qparams": { "group_size": 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": 0.12071378529071808, "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": 0.14198756217956543, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13099680840969086, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12872271239757538, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11404195427894592, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10801760107278824, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07212740927934647, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06251433491706848, "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": 0.060654252767562866, "qparams": { "group_size": 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": 0.060208458453416824, "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": 0.03597613424062729, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030828185379505157, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030682317912578583, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028166934847831726, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027879660949110985, "qparams": { "group_size": 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": 0.01851317659020424, "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": 0.018089789897203445, "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": 0.017820557579398155, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011173505336046219, "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": 0.11404195427894592, "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": 0.11404195427894592, "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": 0.23930390179157257, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20240569114685059, "qparams": { "group_size": 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": 0.1888890415430069, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1542237102985382, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10931836813688278, "qparams": { "group_size": 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": 0.09554090350866318, "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": 0.13122506439685822, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1202835887670517, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11420413106679916, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08351141214370728, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07708428800106049, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06738085299730301, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05809465050697327, "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": 0.05311470851302147, "qparams": { "group_size": 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": 0.05189197137951851, "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": 0.03399006277322769, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.028496503829956055, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028107572346925735, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02362225018441677, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022802084684371948, "qparams": { "group_size": 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": 0.0188304390758276, "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": 0.019375350326299667, "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": 0.017278160899877548, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014360612258315086, "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": 0.1202835887670517, "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": 0.1202835887670517, "qparams": { "group_size": 32, "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": 0.20207421481609344, "qparams": { "group_size": 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": 0.18956758081912994, "qparams": { "group_size": 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": 0.18566320836544037, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.16820116341114044, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09498757869005203, "qparams": { "group_size": 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": 0.09057736396789551, "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": 0.10547754913568497, "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": 0.09763415902853012, "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": 0.09633206576108932, "qparams": { "group_size": 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": 0.08542586863040924, "qparams": { "group_size": 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": 0.0808473527431488, "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": 0.053678106516599655, "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": 0.0467485710978508, "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": 0.045548416674137115, "qparams": { "group_size": 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": 0.04526474326848984, "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": 0.02681359276175499, "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": 0.023534758016467094, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023453228175640106, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021608546376228333, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02142849564552307, "qparams": { "group_size": 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": 0.014078575186431408, "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": 0.01435912773013115, "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": 0.013683576136827469, "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": 0.009707948192954063, "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": 0.10547754913568497, "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": 0.10547754913568497, "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": 0.25993749499320984, "qparams": { "group_size": 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": 0.24406857788562775, "qparams": { "group_size": 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": 0.23918516933918, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21692611277103424, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12211690843105316, "qparams": { "group_size": 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": 0.11653333902359009, "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": 0.1356021761894226, "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": 0.12537631392478943, "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": 0.12381644546985626, "qparams": { "group_size": 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": 0.10992631316184998, "qparams": { "group_size": 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": 0.10423311591148376, "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": 0.06892489641904831, "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": 0.05987607315182686, "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": 0.058394405990839005, "qparams": { "group_size": 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": 0.05804968252778053, "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": 0.03445887938141823, "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": 0.029770145192742348, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0296628475189209, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02724694088101387, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027022799476981163, "qparams": { "group_size": 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": 0.017936047166585922, "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": 0.017524486407637596, "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": 0.017444930970668793, "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": 0.011045075953006744, "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": 0.12211690843105316, "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": 0.12211690843105316, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.26083701848983765, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.2325427532196045, "qparams": { "group_size": 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": 0.221750408411026, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19766055047512054, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11928435415029526, "qparams": { "group_size": 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": 0.10843650996685028, "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": 0.1401265561580658, "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": 0.12891735136508942, "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": 0.12344665080308914, "qparams": { "group_size": 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": 0.1029890924692154, "qparams": { "group_size": 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": 0.09703220427036285, "qparams": { "group_size": 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": 0.0712517499923706, "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": 0.0616760216653347, "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": 0.0573284812271595, "qparams": { "group_size": 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": 0.056256718933582306, "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": 0.03565666079521179, "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": 0.02984781004488468, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029624804854393005, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02661818452179432, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025936992838978767, "qparams": { "group_size": 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": 0.01895960234105587, "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": 0.019019344821572304, "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": 0.017544090747833252, "qparams": { "group_size": 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": 0.01294626947492361, "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": 0.11928435415029526, "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": 0.11928435415029526, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.23.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.13067832589149475, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12200542539358139, "qparams": { "group_size": 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": 0.11849953979253769, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10706979036331177, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06103547289967537, "qparams": { "group_size": 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": 0.05761798471212387, "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": 0.06942083686590195, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06427516788244247, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.061957526952028275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05470946431159973, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05181827396154404, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.035250257700681686, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.030726071447134018, "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": 0.02924049086868763, "qparams": { "group_size": 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": 0.02888040989637375, "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": 0.017625920474529266, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01500649657100439, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014858278445899487, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013721683993935585, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013494626618921757, "qparams": { "group_size": 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": 0.009200597181916237, "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": 0.009138863533735275, "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": 0.00872033927589655, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005913258995860815, "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": 0.12200542539358139, "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": 0.12200542539358139, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.1107102558016777, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10340766608715057, "qparams": { "group_size": 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": 0.1000174805521965, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09036368876695633, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05166608840227127, "qparams": { "group_size": 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": 0.04858045652508736, "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": 0.05953985080122948, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.055050451308488846, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.052454981952905655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04632176086306572, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04398765787482262, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03019697591662407, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.026274750009179115, "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": 0.024747297167778015, "qparams": { "group_size": 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": 0.02438092604279518, "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": 0.015098106116056442, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012717558071017265, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012551666237413883, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011635659262537956, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011403542011976242, "qparams": { "group_size": 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": 0.007877539843320847, "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": 0.007807259447872639, "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": 0.007389286067336798, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0050420211628079414, "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": 0.1107102558016777, "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": 0.1107102558016777, "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": 0.27094343304634094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.25363820791244507, "qparams": { "group_size": 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": 0.2479727864265442, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.22447656095027924, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.127216637134552, "qparams": { "group_size": 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": 0.1209387257695198, "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": 0.14274126291275024, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1315014660358429, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1290794461965561, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11418803781270981, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10817777365446091, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07250487059354782, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06278231739997864, "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": 0.06082814186811447, "qparams": { "group_size": 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": 0.060364726930856705, "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": 0.036148801445961, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030917484313249588, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030759623274207115, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02821299061179161, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027917370200157166, "qparams": { "group_size": 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": 0.018597589805722237, "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": 0.018155382946133614, "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": 0.0178513340651989, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011191303841769695, "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": 0.11418803781270981, "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": 0.11418803781270981, "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": 0.23044326901435852, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18909668922424316, "qparams": { "group_size": 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": 0.17395532131195068, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15257255733013153, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10522028058767319, "qparams": { "group_size": 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": 0.08741830289363861, "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": 0.1258227676153183, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11585129052400589, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11011213809251785, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08244332671165466, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07604009658098221, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06474414467811584, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.056007590144872665, "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": 0.05117683857679367, "qparams": { "group_size": 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": 0.049992628395557404, "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": 0.03264109417796135, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.027580052614212036, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027213189750909805, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023311566561460495, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022525085136294365, "qparams": { "group_size": 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": 0.018030377104878426, "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": 0.018899543210864067, "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": 0.016549386084079742, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014149736613035202, "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": 0.11585129052400589, "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": 0.11585129052400589, "qparams": { "group_size": 32, "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": 0.2088010013103485, "qparams": { "group_size": 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": 0.19578854739665985, "qparams": { "group_size": 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": 0.1917499452829361, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.17379361391067505, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09815508872270584, "qparams": { "group_size": 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": 0.09358977526426315, "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": 0.10884332656860352, "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": 0.10084307193756104, "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": 0.09951922297477722, "qparams": { "group_size": 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": 0.08824906498193741, "qparams": { "group_size": 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": 0.08353133499622345, "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": 0.05543861538171768, "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": 0.04827302694320679, "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": 0.04705202206969261, "qparams": { "group_size": 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": 0.04676263406872749, "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": 0.027690760791301727, "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": 0.02429649792611599, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02421540394425392, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02229977957904339, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022122181951999664, "qparams": { "group_size": 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": 0.014545511454343796, "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": 0.01479980070143938, "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": 0.01414633821696043, "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": 0.009981891140341759, "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": 0.10884332656860352, "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": 0.10884332656860352, "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": 0.26313844323158264, "qparams": { "group_size": 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": 0.24689844250679016, "qparams": { "group_size": 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": 0.2419632077217102, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21932019293308258, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12361442297697067, "qparams": { "group_size": 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": 0.11794338375329971, "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": 0.13696205615997314, "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": 0.12692944705486298, "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": 0.1253284215927124, "qparams": { "group_size": 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": 0.1112150028347969, "qparams": { "group_size": 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": 0.10529331117868423, "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": 0.06960875540971756, "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": 0.06058394908905029, "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": 0.05909565091133118, "qparams": { "group_size": 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": 0.05874832719564438, "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": 0.03473421558737755, "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": 0.03004252351820469, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029943808913230896, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027475163340568542, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027250809594988823, "qparams": { "group_size": 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": 0.01793304830789566, "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": 0.017578475177288055, "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": 0.017433365806937218, "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": 0.010912427678704262, "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": 0.1112150028347969, "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": 0.1112150028347969, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.25801321864128113, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.2303575873374939, "qparams": { "group_size": 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": 0.2194657325744629, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19584551453590393, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11800935864448547, "qparams": { "group_size": 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": 0.1072535589337349, "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": 0.13921822607517242, "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": 0.12808963656425476, "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": 0.12212850898504257, "qparams": { "group_size": 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": 0.1021212637424469, "qparams": { "group_size": 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": 0.09624899923801422, "qparams": { "group_size": 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": 0.07070159912109375, "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": 0.061255618929862976, "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": 0.05673142895102501, "qparams": { "group_size": 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": 0.05562550202012062, "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": 0.035593096166849136, "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": 0.029540736228227615, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029286982491612434, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026408089324831963, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02568802237510681, "qparams": { "group_size": 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": 0.019229263067245483, "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": 0.018871909007430077, "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": 0.017812412232160568, "qparams": { "group_size": 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": 0.012817719019949436, "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": 0.12212850898504257, "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": 0.12212850898504257, "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": 0.13209077715873718, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12348287552595139, "qparams": { "group_size": 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": 0.11992543935775757, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10843086242675781, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06176615506410599, "qparams": { "group_size": 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": 0.05830052122473717, "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": 0.0706094354391098, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06512545794248581, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06270095705986023, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05542076379060745, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05261602997779846, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0358477346599102, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03113020770251751, "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": 0.02960914745926857, "qparams": { "group_size": 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": 0.029246564954519272, "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": 0.017963659018278122, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.015225891955196857, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015067036263644695, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013930676504969597, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013700572773814201, "qparams": { "group_size": 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": 0.00942237488925457, "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": 0.009300898760557175, "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": 0.008915787562727928, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006052036304026842, "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": 0.11992543935775757, "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": 0.11992543935775757, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.11435338854789734, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10683850944042206, "qparams": { "group_size": 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": 0.1035461500287056, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09358701854944229, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05343690514564514, "qparams": { "group_size": 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": 0.050314418971538544, "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": 0.06123281270265579, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05661952495574951, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.054230015724897385, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0479179322719574, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.045470044016838074, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03106260858476162, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.027031470090150833, "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": 0.025583583861589432, "qparams": { "group_size": 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": 0.025236550718545914, "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": 0.015517883002758026, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013101349584758282, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012951168231666088, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011975867673754692, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011754671111702919, "qparams": { "group_size": 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": 0.00806350912898779, "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": 0.007954759523272514, "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": 0.007598054129630327, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005066646263003349, "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": 0.11435338854789734, "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": 0.11435338854789734, "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": 0.26924946904182434, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.25219565629959106, "qparams": { "group_size": 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": 0.24656732380390167, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.22322507202625275, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12649473547935486, "qparams": { "group_size": 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": 0.12025941163301468, "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": 0.14175701141357422, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13073772192001343, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12833108007907867, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11356446146965027, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.1075717955827713, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07202465832233429, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06242552399635315, "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": 0.06046977639198303, "qparams": { "group_size": 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": 0.05999770015478134, "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": 0.03592110425233841, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030729271471500397, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030570819973945618, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028048496693372726, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027751820161938667, "qparams": { "group_size": 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": 0.018479909747838974, "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": 0.0180372204631567, "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": 0.017746860161423683, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011126662604510784, "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": 0.11356446146965027, "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": 0.11356446146965027, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.24.self_attn.o_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.23219527304172516, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.20044666528701782, "qparams": { "group_size": 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": 0.1908266395330429, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15415646135807037, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10405320674180984, "qparams": { "group_size": 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": 0.09317149221897125, "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": 0.12385778874158859, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1114940419793129, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10811839252710342, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0840187519788742, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07587140798568726, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06322773545980453, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05495904013514519, "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": 0.05169428512454033, "qparams": { "group_size": 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": 0.05087990313768387, "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": 0.033054254949092865, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.029424935579299927, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02923242747783661, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02592514455318451, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025459541007876396, "qparams": { "group_size": 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": 0.020134413614869118, "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": 0.021597886458039284, "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": 0.019186556339263916, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.017945080995559692, "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": 0.1114940419793129, "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": 0.1114940419793129, "qparams": { "group_size": 32, "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": 0.21745629608631134, "qparams": { "group_size": 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": 0.20395591855049133, "qparams": { "group_size": 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": 0.1996987909078598, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18099072575569153, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10224489122629166, "qparams": { "group_size": 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": 0.09745485335588455, "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": 0.11345033347606659, "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": 0.10506884753704071, "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": 0.10368253290653229, "qparams": { "group_size": 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": 0.09188506007194519, "qparams": { "group_size": 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": 0.08697400242090225, "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": 0.057713400572538376, "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": 0.050263918936252594, "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": 0.04898253083229065, "qparams": { "group_size": 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": 0.04869204759597778, "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": 0.028847001492977142, "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": 0.025210073217749596, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025127530097961426, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02311536855995655, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022922340780496597, "qparams": { "group_size": 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": 0.01511057373136282, "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": 0.015230708755552769, "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": 0.014692945405840874, "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": 0.010117854923009872, "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": 0.11345033347606659, "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": 0.11345033347606659, "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": 0.2676107585430145, "qparams": { "group_size": 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": 0.2511650025844574, "qparams": { "group_size": 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": 0.24611720442771912, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.22317129373550415, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12586654722690582, "qparams": { "group_size": 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": 0.12004672735929489, "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": 0.13960866630077362, "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": 0.1292218267917633, "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": 0.12763474881649017, "qparams": { "group_size": 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": 0.11322600394487381, "qparams": { "group_size": 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": 0.10723579674959183, "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": 0.07104229182004929, "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": 0.061734650284051895, "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": 0.06021454185247421, "qparams": { "group_size": 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": 0.059857819229364395, "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": 0.03551957383751869, "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": 0.030704716220498085, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030611461028456688, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02811064012348652, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027879813686013222, "qparams": { "group_size": 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": 0.018512364476919174, "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": 0.01811264455318451, "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": 0.018009111285209656, "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": 0.01145433820784092, "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": 0.11322600394487381, "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": 0.11322600394487381, "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": 0.2619823217391968, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23497602343559265, "qparams": { "group_size": 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": 0.22437813878059387, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20048925280570984, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11996804177761078, "qparams": { "group_size": 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": 0.10949810594320297, "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": 0.14065860211849213, "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": 0.12988808751106262, "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": 0.12403050810098648, "qparams": { "group_size": 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": 0.10435152053833008, "qparams": { "group_size": 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": 0.09829186648130417, "qparams": { "group_size": 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": 0.07148870080709457, "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": 0.062085993587970734, "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": 0.05759574845433235, "qparams": { "group_size": 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": 0.05650811269879341, "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": 0.03581956773996353, "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": 0.029836298897862434, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029582282528281212, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026722047477960587, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02601543813943863, "qparams": { "group_size": 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": 0.018986854702234268, "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": 0.018824562430381775, "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": 0.01754874736070633, "qparams": { "group_size": 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": 0.012523054145276546, "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": 0.11996804177761078, "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": 0.11996804177761078, "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": 0.13541877269744873, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.126553013920784, "qparams": { "group_size": 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": 0.12290887534618378, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11115622520446777, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06331712752580643, "qparams": { "group_size": 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": 0.059811070561409, "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": 0.07207520306110382, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06662415713071823, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0642818808555603, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.056768838316202164, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.053819913417100906, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.036568883806467056, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.031823690980672836, "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": 0.03033803403377533, "qparams": { "group_size": 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": 0.029986320063471794, "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": 0.018276091665029526, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.015559328719973564, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01541027519851923, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.014218734577298164, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013993289321660995, "qparams": { "group_size": 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": 0.009529138915240765, "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": 0.00944051519036293, "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": 0.009048563428223133, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0060822670347988605, "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": 0.11115622520446777, "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": 0.11115622520446777, "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": 0.11645641922950745, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10881472378969193, "qparams": { "group_size": 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": 0.10543924570083618, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0953640341758728, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.054455697536468506, "qparams": { "group_size": 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": 0.051275040954351425, "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": 0.06246881186962128, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.057739339768886566, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05529562011361122, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04884905368089676, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04636764898896217, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.031705886125564575, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.027588309720158577, "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": 0.026088010519742966, "qparams": { "group_size": 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": 0.025734443217515945, "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": 0.015841824933886528, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013364823535084724, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013205400668084621, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012213877402245998, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011989151127636433, "qparams": { "group_size": 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": 0.008236614055931568, "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": 0.008130108937621117, "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": 0.007755450904369354, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005179917439818382, "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": 0.11645641922950745, "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": 0.11645641922950745, "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": 0.27697819471359253, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.25952938199043274, "qparams": { "group_size": 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": 0.2537539303302765, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.22981034219264984, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.13022203743457794, "qparams": { "group_size": 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": 0.12382834404706955, "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": 0.1457420438528061, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13439495861530304, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1321273148059845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11687684804201126, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11069600284099579, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07405989617109299, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06417538225650787, "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": 0.06226302310824394, "qparams": { "group_size": 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": 0.06179479509592056, "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": 0.03694484755396843, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03161890059709549, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03147130087018013, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02884889952838421, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028550414368510246, "qparams": { "group_size": 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": 0.018977468833327293, "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": 0.018514646217226982, "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": 0.018243052065372467, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01137601863592863, "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": 0.11687684804201126, "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": 0.11687684804201126, "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": 0.2308996319770813, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19670316576957703, "qparams": { "group_size": 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": 0.1829654723405838, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15092632174491882, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10587626695632935, "qparams": { "group_size": 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": 0.09183476865291595, "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": 0.12829405069351196, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11771119385957718, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11030536890029907, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08294733613729477, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07659217715263367, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06603153049945831, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.056852132081985474, "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": 0.051445793360471725, "qparams": { "group_size": 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": 0.050082772970199585, "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": 0.03327752277255058, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.027572691440582275, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027097433805465698, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023246292024850845, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02236318588256836, "qparams": { "group_size": 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": 0.018342969939112663, "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": 0.01882339082658291, "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": 0.01665906421840191, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013854192569851875, "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": 0.11771119385957718, "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": 0.11771119385957718, "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": 0.2214018851518631, "qparams": { "group_size": 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": 0.2075498104095459, "qparams": { "group_size": 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": 0.20322905480861664, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18424996733665466, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10412415117025375, "qparams": { "group_size": 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": 0.09926318377256393, "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": 0.11564820259809494, "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": 0.10698594897985458, "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": 0.10559780150651932, "qparams": { "group_size": 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": 0.09358861297369003, "qparams": { "group_size": 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": 0.08868394047021866, "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": 0.0589560940861702, "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": 0.05122905597090721, "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": 0.04992423951625824, "qparams": { "group_size": 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": 0.049625739455223083, "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": 0.029493268579244614, "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": 0.025753984227776527, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025672461837530136, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023624364286661148, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02343142405152321, "qparams": { "group_size": 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": 0.015566344372928143, "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": 0.01565040461719036, "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": 0.015150095336139202, "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": 0.010513890534639359, "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": 0.11564820259809494, "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": 0.11564820259809494, "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": 0.26989665627479553, "qparams": { "group_size": 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": 0.2532506585121155, "qparams": { "group_size": 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": 0.2481204718351364, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.22501611709594727, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12697748839855194, "qparams": { "group_size": 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": 0.12109410762786865, "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": 0.14082348346710205, "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": 0.1303725689649582, "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": 0.12876570224761963, "qparams": { "group_size": 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": 0.11420343816280365, "qparams": { "group_size": 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": 0.10815675556659698, "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": 0.0716337040066719, "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": 0.06229635700583458, "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": 0.06075286492705345, "qparams": { "group_size": 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": 0.060383811593055725, "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": 0.03577735275030136, "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": 0.03097722865641117, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03087293729186058, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028347551822662354, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028110498562455177, "qparams": { "group_size": 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": 0.01861201785504818, "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": 0.01827230490744114, "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": 0.018105044960975647, "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": 0.011552425101399422, "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": 0.11420343816280365, "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": 0.11420343816280365, "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": 0.2626211941242218, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23562994599342346, "qparams": { "group_size": 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": 0.224774569272995, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2011839598417282, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12014085054397583, "qparams": { "group_size": 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": 0.1095895990729332, "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": 0.1414215862751007, "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": 0.13047553598880768, "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": 0.12426265329122543, "qparams": { "group_size": 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": 0.1047288253903389, "qparams": { "group_size": 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": 0.09877520799636841, "qparams": { "group_size": 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": 0.07173149287700653, "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": 0.06229905039072037, "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": 0.057659316807985306, "qparams": { "group_size": 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": 0.05651881918311119, "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": 0.035864152014255524, "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": 0.02981964685022831, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02954735979437828, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026734353974461555, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026005761697888374, "qparams": { "group_size": 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": 0.01885409466922283, "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": 0.018770715221762657, "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": 0.017333198338747025, "qparams": { "group_size": 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": 0.012383912689983845, "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": 0.12014085054397583, "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": 0.12014085054397583, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.26.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1347295343875885, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1258833408355713, "qparams": { "group_size": 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": 0.12224689871072769, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11059591174125671, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06306014955043793, "qparams": { "group_size": 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": 0.0594852939248085, "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": 0.07201312482357025, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06644129008054733, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06401945650577545, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.056527528911828995, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.053708188235759735, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03655818849802017, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03174765408039093, "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": 0.030212655663490295, "qparams": { "group_size": 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": 0.029852908104658127, "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": 0.01828213781118393, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.015482649207115173, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01533081941306591, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.014147760346531868, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013917711563408375, "qparams": { "group_size": 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": 0.009529734030365944, "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": 0.009383984841406345, "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": 0.00903056375682354, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006014558020979166, "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": 0.12224689871072769, "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": 0.12224689871072769, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.11788292229175568, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11017101258039474, "qparams": { "group_size": 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": 0.10675989091396332, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09656429290771484, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.055108051747083664, "qparams": { "group_size": 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": 0.051884908229112625, "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": 0.06349112838506699, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.058401286602020264, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05592721328139305, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04940757155418396, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.047016777098178864, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03216457739472389, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.027897734194993973, "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": 0.026420380920171738, "qparams": { "group_size": 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": 0.026059789583086967, "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": 0.01608450897037983, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013535461388528347, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01337909884750843, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012374131940305233, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012147153727710247, "qparams": { "group_size": 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": 0.008382737636566162, "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": 0.00822546798735857, "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": 0.007909574545919895, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0052450974471867085, "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": 0.11788292229175568, "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": 0.11788292229175568, "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": 0.2780638337135315, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2604540288448334, "qparams": { "group_size": 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": 0.25467434525489807, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.230609729886055, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1308853179216385, "qparams": { "group_size": 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": 0.12444702535867691, "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": 0.14653736352920532, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13511237502098083, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13279594480991364, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11744081974029541, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11125967651605606, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07455997914075851, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06454327702522278, "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": 0.06262470781803131, "qparams": { "group_size": 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": 0.062154047191143036, "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": 0.03720511868596077, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03191830962896347, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0317743718624115, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.029144028201699257, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02885376289486885, "qparams": { "group_size": 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": 0.019204040989279747, "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": 0.018867311999201775, "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": 0.018474802374839783, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011865999549627304, "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": 0.11744081974029541, "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": 0.11744081974029541, "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": 0.21687524020671844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18238694965839386, "qparams": { "group_size": 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": 0.17012138664722443, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14361083507537842, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09672395884990692, "qparams": { "group_size": 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": 0.08323455601930618, "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": 0.11865761876106262, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10797249525785446, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10296319425106049, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07903285324573517, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07229288667440414, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06061652675271034, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05274472385644913, "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": 0.04771260917186737, "qparams": { "group_size": 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": 0.046454984694719315, "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": 0.030749980360269547, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026599470525979996, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026288630440831184, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023435521870851517, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022688621655106544, "qparams": { "group_size": 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": 0.017449868842959404, "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": 0.019270922988653183, "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": 0.015948129817843437, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01537595596164465, "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": 0.11865761876106262, "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": 0.11865761876106262, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "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": 0.22333821654319763, "qparams": { "group_size": 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": 0.20940835773944855, "qparams": { "group_size": 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": 0.20507685840129852, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1857859045267105, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10509465634822845, "qparams": { "group_size": 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": 0.10016694664955139, "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": 0.11663287878036499, "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": 0.10799086838960648, "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": 0.10659419745206833, "qparams": { "group_size": 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": 0.09441237151622772, "qparams": { "group_size": 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": 0.0893600657582283, "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": 0.059402287006378174, "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": 0.051684483885765076, "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": 0.05036666616797447, "qparams": { "group_size": 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": 0.05005284398794174, "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": 0.02966143935918808, "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": 0.025925641879439354, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025841601192951202, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02376011572778225, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023561950773000717, "qparams": { "group_size": 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": 0.01554079633206129, "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": 0.0156736858189106, "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": 0.015113718807697296, "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": 0.010424705222249031, "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": 0.11663287878036499, "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": 0.11663287878036499, "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": 0.2711191177368164, "qparams": { "group_size": 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": 0.2543366253376007, "qparams": { "group_size": 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": 0.2492310106754303, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.22589832544326782, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12749895453453064, "qparams": { "group_size": 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": 0.12159547209739685, "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": 0.14140599966049194, "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": 0.13090519607067108, "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": 0.129268079996109, "qparams": { "group_size": 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": 0.1146274134516716, "qparams": { "group_size": 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": 0.10852230340242386, "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": 0.07178191095590591, "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": 0.06252666562795639, "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": 0.06097632646560669, "qparams": { "group_size": 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": 0.060603342950344086, "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": 0.03586050495505333, "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": 0.03101353719830513, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030907971784472466, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028343576937913895, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028108732774853706, "qparams": { "group_size": 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": 0.018558034673333168, "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": 0.01816549338400364, "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": 0.018043801188468933, "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": 0.011308182030916214, "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": 0.1146274134516716, "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": 0.1146274134516716, "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": 0.2624221742153168, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23595620691776276, "qparams": { "group_size": 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": 0.22530263662338257, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20173530280590057, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12011420726776123, "qparams": { "group_size": 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": 0.10979869216680527, "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": 0.14142395555973053, "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": 0.13048610091209412, "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": 0.12426948547363281, "qparams": { "group_size": 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": 0.10509609431028366, "qparams": { "group_size": 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": 0.09916435182094574, "qparams": { "group_size": 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": 0.07208491116762161, "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": 0.062428124248981476, "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": 0.057688310742378235, "qparams": { "group_size": 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": 0.056539032608270645, "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": 0.03622843325138092, "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": 0.029912959784269333, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029642220586538315, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02691761404275894, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02618398144841194, "qparams": { "group_size": 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": 0.01938127540051937, "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": 0.018948666751384735, "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": 0.017850542441010475, "qparams": { "group_size": 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": 0.012648245319724083, "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": 0.12011420726776123, "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": 0.12011420726776123, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.1454668641090393, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13596691191196442, "qparams": { "group_size": 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": 0.13230356574058533, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11958257108926773, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06817647814750671, "qparams": { "group_size": 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": 0.06445496529340744, "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": 0.07718474417924881, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07141443341970444, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0692112073302269, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.061087917536497116, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05781807377934456, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.039116181433200836, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03410720452666283, "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": 0.0326339490711689, "qparams": { "group_size": 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": 0.03229020908474922, "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": 0.019541652873158455, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.016666358336806297, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.016529854387044907, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01521090418100357, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014987030066549778, "qparams": { "group_size": 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": 0.010120564140379429, "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": 0.009994784370064735, "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": 0.009635942056775093, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006309958640486002, "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": 0.11958257108926773, "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": 0.11958257108926773, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.27.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.12551528215408325, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11733756214380264, "qparams": { "group_size": 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": 0.11395540833473206, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10301994532346725, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.058821164071559906, "qparams": { "group_size": 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": 0.055550217628479004, "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": 0.06707216799259186, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.061932481825351715, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05971523001790047, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0527234822511673, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05005263537168503, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03404434397816658, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029575306922197342, "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": 0.028179548680782318, "qparams": { "group_size": 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": 0.027839576825499535, "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": 0.017018387094140053, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014406170696020126, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014268171973526478, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013153690844774246, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01294309925287962, "qparams": { "group_size": 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": 0.008829339407384396, "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": 0.008678612299263477, "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": 0.008370785973966122, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005495382007211447, "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": 0.11733756214380264, "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": 0.11733756214380264, "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": 0.28290170431137085, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2648734748363495, "qparams": { "group_size": 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": 0.25920215249061584, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.23458713293075562, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.13310614228248596, "qparams": { "group_size": 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": 0.1265919804573059, "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": 0.14872196316719055, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13721051812171936, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13502071797847748, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11935541778802872, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.1130785346031189, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07561182230710983, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06553581357002258, "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": 0.06366977095603943, "qparams": { "group_size": 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": 0.06320954114198685, "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": 0.037722550332546234, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.032347504049539566, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03221122920513153, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02950141578912735, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029214652255177498, "qparams": { "group_size": 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": 0.019413495436310768, "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": 0.01895127072930336, "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": 0.01870892569422722, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01170633640140295, "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": 0.11935541778802872, "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": 0.11935541778802872, "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": 0.2245260328054428, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19548264145851135, "qparams": { "group_size": 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": 0.1846010386943817, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1573038548231125, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10266254842281342, "qparams": { "group_size": 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": 0.09154476225376129, "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": 0.12208559364080429, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1125902310013771, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10679138451814651, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08379627764225006, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07766196131706238, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06262048333883286, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05408778041601181, "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": 0.04947558045387268, "qparams": { "group_size": 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": 0.048340439796447754, "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": 0.03142167255282402, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025856344029307365, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025471633300185204, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02213481441140175, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021370019763708115, "qparams": { "group_size": 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": 0.01685953699052334, "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": 0.01676221378147602, "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": 0.015370959416031837, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01147421170026064, "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": 0.12208559364080429, "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": 0.12208559364080429, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.27.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.22716976702213287, "qparams": { "group_size": 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": 0.21296162903308868, "qparams": { "group_size": 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": 0.20858550071716309, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1888861209154129, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10696892440319061, "qparams": { "group_size": 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": 0.10192114114761353, "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": 0.1188310906291008, "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": 0.10996732860803604, "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": 0.10852924734354019, "qparams": { "group_size": 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": 0.09603272378444672, "qparams": { "group_size": 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": 0.09088417887687683, "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": 0.060517825186252594, "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": 0.05261842906475067, "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": 0.051266323775053024, "qparams": { "group_size": 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": 0.050945691764354706, "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": 0.03025549277663231, "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": 0.02636655978858471, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02627679333090782, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.024146607145667076, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02394591085612774, "qparams": { "group_size": 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": 0.015892837196588516, "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": 0.015906311571598053, "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": 0.01544522400945425, "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": 0.010533902794122696, "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": 0.1188310906291008, "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": 0.1188310906291008, "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": 0.27420681715011597, "qparams": { "group_size": 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": 0.25718095898628235, "qparams": { "group_size": 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": 0.25201576948165894, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2283465564250946, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12907548248767853, "qparams": { "group_size": 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": 0.12304611504077911, "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": 0.14309810101985931, "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": 0.13257445394992828, "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": 0.1308644860982895, "qparams": { "group_size": 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": 0.11598342657089233, "qparams": { "group_size": 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": 0.10974625498056412, "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": 0.07276982069015503, "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": 0.0633181482553482, "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": 0.061725616455078125, "qparams": { "group_size": 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": 0.061347730457782745, "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": 0.03630413860082626, "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": 0.03138856589794159, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.031284090131521225, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02867116406559944, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02842605859041214, "qparams": { "group_size": 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": 0.018756480887532234, "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": 0.01837346702814102, "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": 0.018230339512228966, "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": 0.01140796672552824, "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": 0.11598342657089233, "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": 0.11598342657089233, "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": 0.26610398292541504, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23999308049678802, "qparams": { "group_size": 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": 0.22930379211902618, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2053898274898529, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12186470627784729, "qparams": { "group_size": 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": 0.11170096695423126, "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": 0.1434285044670105, "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": 0.1323135793209076, "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": 0.12594901025295258, "qparams": { "group_size": 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": 0.10684289038181305, "qparams": { "group_size": 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": 0.10077009350061417, "qparams": { "group_size": 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": 0.07276375591754913, "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": 0.06313492357730865, "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": 0.05843265354633331, "qparams": { "group_size": 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": 0.05728727951645851, "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": 0.03642905503511429, "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": 0.030098164454102516, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02980857901275158, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02707003243267536, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02632186748087406, "qparams": { "group_size": 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": 0.01918252930045128, "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": 0.018782559782266617, "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": 0.017648771405220032, "qparams": { "group_size": 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": 0.012154822237789631, "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": 0.12186470627784729, "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": 0.12186470627784729, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.28.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.14414554834365845, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13467514514923096, "qparams": { "group_size": 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": 0.13079369068145752, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11813019961118698, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06759880483150482, "qparams": { "group_size": 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": 0.06376951187849045, "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": 0.07685772329568863, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0711190328001976, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06863138824701309, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06051063910126686, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05727922543883324, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03904305025935173, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03402739390730858, "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": 0.03241050988435745, "qparams": { "group_size": 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": 0.03202369809150696, "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": 0.01953456923365593, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.0166226364672184, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01646566390991211, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015175643377006054, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014931253157556057, "qparams": { "group_size": 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": 0.010202317498624325, "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": 0.010104265995323658, "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": 0.009675397537648678, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006518927402794361, "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": 0.11813019961118698, "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": 0.11813019961118698, "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": 0.12435005605220795, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11615114659070969, "qparams": { "group_size": 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": 0.11268840730190277, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10182392597198486, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.058245208114385605, "qparams": { "group_size": 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": 0.05486968159675598, "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": 0.06665335595607758, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.061575859785079956, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.059140127152204514, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05213836953043938, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04950675740838051, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.033827729523181915, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029404299333691597, "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": 0.02790418080985546, "qparams": { "group_size": 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": 0.027545105665922165, "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": 0.016903722658753395, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014280701987445354, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014128846116364002, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01303150039166212, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012800447642803192, "qparams": { "group_size": 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": 0.008788627572357655, "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": 0.008644402958452702, "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": 0.008293970488011837, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00548928789794445, "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": 0.11615114659070969, "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": 0.11615114659070969, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.27844634652137756, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.26049691438674927, "qparams": { "group_size": 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": 0.25463396310806274, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.23032473027706146, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1308877021074295, "qparams": { "group_size": 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": 0.12432469427585602, "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": 0.1465698629617691, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1351642608642578, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13282662630081177, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11726520210504532, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11096726357936859, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07449407875537872, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0645575150847435, "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": 0.06258663535118103, "qparams": { "group_size": 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": 0.06211595609784126, "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": 0.0371394120156765, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03176620602607727, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03161434084177017, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028931034728884697, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028627028688788414, "qparams": { "group_size": 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": 0.019070975482463837, "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": 0.018590278923511505, "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": 0.018332213163375854, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011388882994651794, "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": 0.11726520210504532, "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": 0.11726520210504532, "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": 0.20683524012565613, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.16997811198234558, "qparams": { "group_size": 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": 0.1545385718345642, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12434930354356766, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09040902554988861, "qparams": { "group_size": 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": 0.07448854297399521, "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": 0.11707080900669098, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10515279322862625, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.09641869366168976, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06991501152515411, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06568160653114319, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05924372375011444, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.051414020359516144, "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": 0.04481933265924454, "qparams": { "group_size": 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": 0.04312477260828018, "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": 0.030469056218862534, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025264371186494827, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02472524717450142, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021618293598294258, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020580193027853966, "qparams": { "group_size": 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": 0.0176565982401371, "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": 0.01877354271709919, "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": 0.015770450234413147, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.014929966069757938, "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": 0.11707080900669098, "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": 0.11707080900669098, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "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": 0.23261266946792603, "qparams": { "group_size": 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": 0.21787603199481964, "qparams": { "group_size": 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": 0.21333566308021545, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1932246834039688, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10965342819690704, "qparams": { "group_size": 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": 0.10441314429044724, "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": 0.12192010879516602, "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": 0.11274363845586777, "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": 0.11125805974006653, "qparams": { "group_size": 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": 0.09840629249811172, "qparams": { "group_size": 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": 0.09310551732778549, "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": 0.06215311214327812, "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": 0.05401833355426788, "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": 0.05260564386844635, "qparams": { "group_size": 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": 0.05227562412619591, "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": 0.03115777298808098, "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": 0.027143772691488266, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02705109491944313, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.024863919243216515, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02465672791004181, "qparams": { "group_size": 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": 0.016559703275561333, "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": 0.016501640900969505, "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": 0.01610661670565605, "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": 0.011082887649536133, "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": 0.12192010879516602, "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": 0.12192010879516602, "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": 0.2782207131385803, "qparams": { "group_size": 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": 0.26076140999794006, "qparams": { "group_size": 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": 0.2554476857185364, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.23150889575481415, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.13100796937942505, "qparams": { "group_size": 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": 0.1248408779501915, "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": 0.1456390917301178, "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": 0.13462914526462555, "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": 0.13289479911327362, "qparams": { "group_size": 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": 0.117650605738163, "qparams": { "group_size": 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": 0.1113661527633667, "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": 0.07404039800167084, "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": 0.0643686056137085, "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": 0.06271992623806, "qparams": { "group_size": 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": 0.06233938783407211, "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": 0.037070322781801224, "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": 0.03199825435876846, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03188787400722504, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02923235297203064, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02898743748664856, "qparams": { "group_size": 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": 0.019439471885561943, "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": 0.018887152895331383, "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": 0.01891118474304676, "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": 0.011952940374612808, "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": 0.117650605738163, "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": 0.117650605738163, "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": 0.2673376202583313, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24180130660533905, "qparams": { "group_size": 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": 0.23087701201438904, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20696505904197693, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1226993203163147, "qparams": { "group_size": 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": 0.11250478774309158, "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": 0.14476275444030762, "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": 0.1337203085422516, "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": 0.1267210841178894, "qparams": { "group_size": 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": 0.10793774574995041, "qparams": { "group_size": 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": 0.10189246386289597, "qparams": { "group_size": 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": 0.07358329743146896, "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": 0.0639977753162384, "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": 0.058971792459487915, "qparams": { "group_size": 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": 0.05775155872106552, "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": 0.036866724491119385, "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": 0.030647337436676025, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030306190252304077, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027703121304512024, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026926560327410698, "qparams": { "group_size": 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": 0.019415894523262978, "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": 0.019522907212376595, "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": 0.017768755555152893, "qparams": { "group_size": 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": 0.01311066746711731, "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": 0.1226993203163147, "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": 0.1226993203163147, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.14894142746925354, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13920268416404724, "qparams": { "group_size": 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": 0.13551558554172516, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12248169630765915, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06992179155349731, "qparams": { "group_size": 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": 0.06610281765460968, "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": 0.07906170934438705, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07315567135810852, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07099087536334991, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06263535469770432, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05924845114350319, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.040085569024086, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03493626043200493, "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": 0.03346247225999832, "qparams": { "group_size": 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": 0.03311719372868538, "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": 0.020022934302687645, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.017069388180971146, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.016927866265177727, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015560268424451351, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015331976115703583, "qparams": { "group_size": 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": 0.010343835689127445, "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": 0.010181124322116375, "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": 0.009858601726591587, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006363655440509319, "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": 0.12248169630765915, "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": 0.12248169630765915, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.29.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.12969988584518433, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12117781490087509, "qparams": { "group_size": 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": 0.11778715997934341, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10643968731164932, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.060838956385850906, "qparams": { "group_size": 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": 0.057462938129901886, "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": 0.06918951869010925, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06388701498508453, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06175227463245392, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05449346452951431, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0516376867890358, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03511219471693039, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.030509015545248985, "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": 0.029135961085557938, "qparams": { "group_size": 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": 0.028805876150727272, "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": 0.01753934845328331, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014874966815114021, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014742951840162277, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01356732938438654, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013358767144382, "qparams": { "group_size": 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": 0.009087440557777882, "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": 0.008919354528188705, "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": 0.008640267886221409, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005614707246422768, "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": 0.12117781490087509, "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": 0.12117781490087509, "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": 0.2866753935813904, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.26841971278190613, "qparams": { "group_size": 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": 0.26249054074287415, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.23750510811805725, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1349536031484604, "qparams": { "group_size": 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": 0.12828576564788818, "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": 0.15079091489315033, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13919121026992798, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13695624470710754, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.12094501405954361, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11448881775140762, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07662615925073624, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06646973639726639, "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": 0.06455834954977036, "qparams": { "group_size": 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": 0.06409577280282974, "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": 0.038251861929893494, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.032759860157966614, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03261521831154823, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.029832249507308006, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029549820348620415, "qparams": { "group_size": 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": 0.01964513026177883, "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": 0.01911577209830284, "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": 0.018924767151474953, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011682996526360512, "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": 0.12094501405954361, "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": 0.12094501405954361, "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": 0.22971272468566895, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1875782310962677, "qparams": { "group_size": 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": 0.1702498495578766, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14846192300319672, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10170173645019531, "qparams": { "group_size": 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": 0.08559592813253403, "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": 0.12873609364032745, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11847249418497086, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10996192693710327, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08152895420789719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07772678881883621, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06620920449495316, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05714584141969681, "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": 0.04949720948934555, "qparams": { "group_size": 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": 0.047540340572595596, "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": 0.033293578773736954, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026503007858991623, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025923382490873337, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02264741249382496, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021339092403650284, "qparams": { "group_size": 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": 0.018229065462946892, "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": 0.018465440720319748, "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": 0.015883933752775192, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013249844312667847, "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": 0.11847249418497086, "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": 0.11847249418497086, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "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": 0.23437000811100006, "qparams": { "group_size": 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": 0.21952342987060547, "qparams": { "group_size": 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": 0.21491189301013947, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19455008208751678, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1104574128985405, "qparams": { "group_size": 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": 0.10515255481004715, "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": 0.12292974442243576, "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": 0.11358734220266342, "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": 0.11204946786165237, "qparams": { "group_size": 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": 0.09906992316246033, "qparams": { "group_size": 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": 0.09379860013723373, "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": 0.06264782696962357, "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": 0.05439886823296547, "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": 0.052977945655584335, "qparams": { "group_size": 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": 0.05263636261224747, "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": 0.03134937956929207, "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": 0.02730727754533291, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02721281349658966, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02499639429152012, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.024785548448562622, "qparams": { "group_size": 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": 0.01656348817050457, "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": 0.016560375690460205, "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": 0.016101446002721786, "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": 0.011073017492890358, "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": 0.11358734220266342, "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": 0.11358734220266342, "qparams": { "group_size": 32, "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": 0.2801651358604431, "qparams": { "group_size": 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": 0.26265084743499756, "qparams": { "group_size": 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": 0.2572595179080963, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2328353226184845, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.13197430968284607, "qparams": { "group_size": 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": 0.1256929636001587, "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": 0.146398663520813, "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": 0.13561706244945526, "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": 0.1338612288236618, "qparams": { "group_size": 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": 0.1183662861585617, "qparams": { "group_size": 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": 0.11195848882198334, "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": 0.07445360720157623, "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": 0.0648060142993927, "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": 0.06313967704772949, "qparams": { "group_size": 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": 0.06275148689746857, "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": 0.037181541323661804, "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": 0.03210357204079628, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03199395164847374, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.029297111555933952, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02904154732823372, "qparams": { "group_size": 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": 0.019242802634835243, "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": 0.018810445442795753, "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": 0.01869155466556549, "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": 0.011685309000313282, "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": 0.1183662861585617, "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": 0.1183662861585617, "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": 0.26934799551963806, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24315670132637024, "qparams": { "group_size": 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": 0.2317056655883789, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2078189253807068, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12342742830514908, "qparams": { "group_size": 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": 0.11291220784187317, "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": 0.1471547931432724, "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": 0.13531604409217834, "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": 0.12769174575805664, "qparams": { "group_size": 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": 0.10866058617830276, "qparams": { "group_size": 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": 0.10276050865650177, "qparams": { "group_size": 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": 0.0749036893248558, "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": 0.0648195818066597, "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": 0.05941977724432945, "qparams": { "group_size": 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": 0.05811157077550888, "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": 0.037851206958293915, "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": 0.03104563057422638, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030677858740091324, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028126470744609833, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02728940360248089, "qparams": { "group_size": 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": 0.020597387105226517, "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": 0.020045680925250053, "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": 0.018860677257180214, "qparams": { "group_size": 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": 0.013690019957721233, "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": 0.10866058617830276, "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": 0.10866058617830276, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.1538248211145401, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14378181099891663, "qparams": { "group_size": 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": 0.13988855481147766, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.126469686627388, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07227218896150589, "qparams": { "group_size": 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": 0.06833384186029434, "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": 0.08186361938714981, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07559770345687866, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.073370061814785, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06466586142778397, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0612567737698555, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04158603772521019, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03615818917751312, "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": 0.03463028743863106, "qparams": { "group_size": 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": 0.03426807373762131, "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": 0.020793883129954338, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01772928051650524, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01758810319006443, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01617014780640602, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015939250588417053, "qparams": { "group_size": 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": 0.010818183422088623, "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": 0.01067607756704092, "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": 0.010302864946424961, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0068309251219034195, "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": 0.08186361938714981, "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": 0.08186361938714981, "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": 0.13363170623779297, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12480899691581726, "qparams": { "group_size": 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": 0.12136760354042053, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10970188677310944, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0627613216638565, "qparams": { "group_size": 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": 0.0592602901160717, "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": 0.07145543396472931, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06587991118431091, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06370499730110168, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05618719384074211, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.053270403295755386, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03629307076334953, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03148827701807022, "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": 0.03006678633391857, "qparams": { "group_size": 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": 0.029726101085543633, "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": 0.018140694126486778, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.015368988737463951, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01523368526250124, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.014024706557393074, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013804817572236061, "qparams": { "group_size": 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": 0.009420707821846008, "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": 0.009251104667782784, "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": 0.008953958749771118, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0058631254360079765, "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": 0.12136760354042053, "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": 0.12136760354042053, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.28875836730003357, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.27030932903289795, "qparams": { "group_size": 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": 0.26440656185150146, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.23918336629867554, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.13613052666187286, "qparams": { "group_size": 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": 0.12936992943286896, "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": 0.15190666913986206, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.14027544856071472, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13806594908237457, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.12190022319555283, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11534720659255981, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07720248401165009, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06697909533977509, "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": 0.0650855079293251, "qparams": { "group_size": 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": 0.06463661044836044, "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": 0.03852732479572296, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03307230398058891, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03292866051197052, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.030122052878141403, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02982480265200138, "qparams": { "group_size": 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": 0.019814740866422653, "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": 0.019362768158316612, "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": 0.019108101725578308, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0119347358122468, "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": 0.12190022319555283, "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": 0.12190022319555283, "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": 0.23939934372901917, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19485969841480255, "qparams": { "group_size": 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": 0.1758331060409546, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15423302352428436, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10809995234012604, "qparams": { "group_size": 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": 0.08766619861125946, "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": 0.135290265083313, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12452992796897888, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11450213938951492, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08507885038852692, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.08001893758773804, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06928765028715134, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05977078899741173, "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": 0.05233916640281677, "qparams": { "group_size": 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": 0.05044684186577797, "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": 0.034691061824560165, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.027699144557118416, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027035951614379883, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023255405947566032, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021960709244012833, "qparams": { "group_size": 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": 0.018557239323854446, "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": 0.01880880817770958, "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": 0.016189521178603172, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013111373409628868, "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": 0.11450213938951492, "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": 0.11450213938951492, "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": 0.23052453994750977, "qparams": { "group_size": 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": 0.2159086912870407, "qparams": { "group_size": 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": 0.21129852533340454, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19132782518863678, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10873452574014664, "qparams": { "group_size": 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": 0.10347022116184235, "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": 0.12076641619205475, "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": 0.11185115575790405, "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": 0.11031337827444077, "qparams": { "group_size": 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": 0.09745319932699203, "qparams": { "group_size": 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": 0.09220386296510696, "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": 0.06153460219502449, "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": 0.05355318635702133, "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": 0.05213598906993866, "qparams": { "group_size": 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": 0.051799383014440536, "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": 0.03076355531811714, "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": 0.026826390996575356, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026734326034784317, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0245344415307045, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02432071417570114, "qparams": { "group_size": 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": 0.01613115519285202, "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": 0.016205161809921265, "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": 0.015668457373976707, "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": 0.010745350271463394, "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": 0.12076641619205475, "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": 0.12076641619205475, "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": 0.280342161655426, "qparams": { "group_size": 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": 0.2626727819442749, "qparams": { "group_size": 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": 0.257264643907547, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.23297563195228577, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1321529746055603, "qparams": { "group_size": 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": 0.1258397400379181, "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": 0.1467130482196808, "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": 0.13587506115436554, "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": 0.13406062126159668, "qparams": { "group_size": 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": 0.1185646802186966, "qparams": { "group_size": 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": 0.11204249411821365, "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": 0.07470224052667618, "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": 0.06493078172206879, "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": 0.06326359510421753, "qparams": { "group_size": 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": 0.06286994367837906, "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": 0.037286415696144104, "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": 0.03224223107099533, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03212873637676239, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.029423780739307404, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02917003259062767, "qparams": { "group_size": 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": 0.019367428496479988, "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": 0.018986498937010765, "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": 0.01881755329668522, "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": 0.011940456926822662, "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": 0.1185646802186966, "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": 0.1185646802186966, "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": 0.2683582007884979, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24152907729148865, "qparams": { "group_size": 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": 0.22991251945495605, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20647335052490234, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12269552797079086, "qparams": { "group_size": 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": 0.1120036169886589, "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": 0.1457538604736328, "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": 0.13483969867229462, "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": 0.12719345092773438, "qparams": { "group_size": 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": 0.10798449069261551, "qparams": { "group_size": 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": 0.10206601768732071, "qparams": { "group_size": 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": 0.0741044133901596, "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": 0.06447894871234894, "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": 0.05898250266909599, "qparams": { "group_size": 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": 0.05763813853263855, "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": 0.03710004314780235, "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": 0.03063441812992096, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030265074223279953, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027698218822479248, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02683732472360134, "qparams": { "group_size": 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": 0.019593210890889168, "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": 0.019574666395783424, "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": 0.01783120259642601, "qparams": { "group_size": 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": 0.013058256357908249, "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": 0.12269552797079086, "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": 0.12269552797079086, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.14854586124420166, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13872413337230682, "qparams": { "group_size": 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": 0.1349111944437027, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12190856039524078, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06976945698261261, "qparams": { "group_size": 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": 0.06587875634431839, "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": 0.07902972400188446, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0731571689248085, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07085951417684555, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06241822987794876, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05899977684020996, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.040068939328193665, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03493892028927803, "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": 0.033424925059080124, "qparams": { "group_size": 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": 0.03306083381175995, "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": 0.020029904320836067, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01707962341606617, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01693572849035263, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015570515766739845, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015334256924688816, "qparams": { "group_size": 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": 0.010397268459200859, "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": 0.010253953747451305, "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": 0.00990103930234909, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00649286201223731, "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": 0.12190856039524078, "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": 0.12190856039524078, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.31.self_attn.k_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.12858279049396515, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12000343948602676, "qparams": { "group_size": 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": 0.11659799516201019, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10533402860164642, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06034772843122482, "qparams": { "group_size": 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": 0.05692930147051811, "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": 0.06867519021034241, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06343965977430344, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06129619479179382, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05397848039865494, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05114709958434105, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.034874700009822845, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.030322063714265823, "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": 0.028920788317918777, "qparams": { "group_size": 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": 0.028585176914930344, "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": 0.01744052767753601, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014797305688261986, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014659612439572811, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013492073863744736, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01328087504953146, "qparams": { "group_size": 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": 0.009064646437764168, "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": 0.008927654474973679, "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": 0.008609212003648281, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005684921983629465, "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": 0.12000343948602676, "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": 0.12000343948602676, "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": 0.29015398025512695, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2713492214679718, "qparams": { "group_size": 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": 0.26547884941101074, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.24003714323043823, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.13665835559368134, "qparams": { "group_size": 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": 0.12986285984516144, "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": 0.15259511768817902, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1408478021621704, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13869042694568634, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.12234514951705933, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11566632241010666, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0775194838643074, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06729110330343246, "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": 0.06536239385604858, "qparams": { "group_size": 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": 0.06490571796894073, "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": 0.03868560492992401, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03317282348871231, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.033027201890945435, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.03018096648156643, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029883647337555885, "qparams": { "group_size": 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": 0.019860751926898956, "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": 0.019348306581377983, "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": 0.019145356491208076, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011819291859865189, "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": 0.12234514951705933, "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": 0.12234514951705933, "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": 0.2258044183254242, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1825459748506546, "qparams": { "group_size": 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": 0.16304005682468414, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14276909828186035, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10190495103597641, "qparams": { "group_size": 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": 0.08023395389318466, "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": 0.12911635637283325, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.1188608705997467, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1080835685133934, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0791257992386818, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07529158890247345, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06621334701776505, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05718411132693291, "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": 0.04946477711200714, "qparams": { "group_size": 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": 0.04749186709523201, "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": 0.033185847103595734, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026369299739599228, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02566032111644745, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0220277551561594, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02069401741027832, "qparams": { "group_size": 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": 0.0178567785769701, "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": 0.01822645775973797, "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": 0.015403551049530506, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012900219298899174, "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": 0.1188608705997467, "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": 0.1188608705997467, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.31.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.22828802466392517, "qparams": { "group_size": 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": 0.2137569785118103, "qparams": { "group_size": 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": 0.20921513438224792, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18926534056663513, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10774033516645432, "qparams": { "group_size": 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": 0.10251246392726898, "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": 0.11974091827869415, "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": 0.11084140092134476, "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": 0.10931894183158875, "qparams": { "group_size": 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": 0.09652720391750336, "qparams": { "group_size": 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": 0.09123197942972183, "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": 0.06104555353522301, "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": 0.05312091112136841, "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": 0.051712516695261, "qparams": { "group_size": 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": 0.05136994644999504, "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": 0.03055676631629467, "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": 0.02667032927274704, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026570064947009087, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.024397168308496475, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02418486215174198, "qparams": { "group_size": 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": 0.016135703772306442, "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": 0.016210153698921204, "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": 0.015672573819756508, "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": 0.010870428755879402, "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": 0.11974091827869415, "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": 0.11974091827869415, "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": 0.2809968590736389, "qparams": { "group_size": 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": 0.26317915320396423, "qparams": { "group_size": 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": 0.2577138841152191, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.23317018151283264, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.13248619437217712, "qparams": { "group_size": 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": 0.1261223405599594, "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": 0.14706861972808838, "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": 0.1361672580242157, "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": 0.1344214379787445, "qparams": { "group_size": 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": 0.11873592436313629, "qparams": { "group_size": 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": 0.1122635006904602, "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": 0.0748300850391388, "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": 0.06509163975715637, "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": 0.06339873373508453, "qparams": { "group_size": 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": 0.06300436705350876, "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": 0.03734554722905159, "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": 0.032260261476039886, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.032147638499736786, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.029409365728497505, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029155656695365906, "qparams": { "group_size": 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": 0.019327519461512566, "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": 0.01893322728574276, "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": 0.018764572218060493, "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": 0.01180023979395628, "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": 0.11873592436313629, "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": 0.11873592436313629, "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": 0.2723458707332611, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.2454710304737091, "qparams": { "group_size": 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": 0.23366233706474304, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20979852974414825, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12470456212759018, "qparams": { "group_size": 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": 0.11391358822584152, "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": 0.14949458837509155, "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": 0.13707217574119568, "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": 0.12930110096931458, "qparams": { "group_size": 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": 0.10983613133430481, "qparams": { "group_size": 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": 0.10386364161968231, "qparams": { "group_size": 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": 0.07566656172275543, "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": 0.06565181910991669, "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": 0.060060419142246246, "qparams": { "group_size": 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": 0.05869070440530777, "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": 0.03817049413919449, "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": 0.031371694058179855, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03099752776324749, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028423914685845375, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0275654885917902, "qparams": { "group_size": 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": 0.02062673680484295, "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": 0.02026715688407421, "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": 0.018858101218938828, "qparams": { "group_size": 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": 0.01382228173315525, "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": 0.10983613133430481, "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": 0.10983613133430481, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.1514318436384201, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14131341874599457, "qparams": { "group_size": 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": 0.1374426931142807, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1240316703915596, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07111839950084686, "qparams": { "group_size": 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": 0.06713298708200455, "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": 0.08063852787017822, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07458590716123581, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07223781943321228, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06352713704109192, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06006466969847679, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.040917426347732544, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03562392666935921, "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": 0.03406552970409393, "qparams": { "group_size": 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": 0.033687856048345566, "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": 0.02043731138110161, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.017412640154361725, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01726296916604042, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015849130228161812, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015611114911735058, "qparams": { "group_size": 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": 0.010589550249278545, "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": 0.0104522081092, "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": 0.010066450573503971, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0066132694482803345, "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": 0.08063852787017822, "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": 0.08063852787017822, "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": 0.1295476108789444, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12089929729700089, "qparams": { "group_size": 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": 0.11734991520643234, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10595070570707321, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06081676483154297, "qparams": { "group_size": 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": 0.057320788502693176, "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": 0.06938499957323074, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06407540291547775, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0617663711309433, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05435322970151901, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05149180069565773, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03524132817983627, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03061419166624546, "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": 0.02915273979306221, "qparams": { "group_size": 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": 0.028793705627322197, "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": 0.017614830285310745, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014918290078639984, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014772210270166397, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013588023371994495, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013365679420530796, "qparams": { "group_size": 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": 0.009153281338512897, "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": 0.009010717272758484, "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": 0.008672177791595459, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005730645265430212, "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": 0.12089929729700089, "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": 0.12089929729700089, "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": 0.29454168677330017, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.27540507912635803, "qparams": { "group_size": 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": 0.26940619945526123, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.24350029230117798, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.13890649378299713, "qparams": { "group_size": 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": 0.13191857933998108, "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": 0.1551733762025833, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.14312157034873962, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.14093486964702606, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.12420956790447235, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11741802841424942, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07890110462903976, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06837772578001022, "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": 0.06641126424074173, "qparams": { "group_size": 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": 0.0659450814127922, "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": 0.03934953734278679, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.033735908567905426, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03359267860651016, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.030690031126141548, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030381083488464355, "qparams": { "group_size": 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": 0.0202198289334774, "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": 0.019739171490073204, "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": 0.019479740411043167, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012142499908804893, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11741802841424942, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, "best_option": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11741802841424942, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "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": 0.23119333386421204, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18746760487556458, "qparams": { "group_size": 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": 0.16898085176944733, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.14399085938930511, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10232767462730408, "qparams": { "group_size": 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": 0.0851232185959816, "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": 0.13064348697662354, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11978105455636978, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11032801866531372, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08104950189590454, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07569217681884766, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06701216101646423, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.057435888797044754, "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": 0.049418967217206955, "qparams": { "group_size": 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": 0.04735346883535385, "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": 0.03352528437972069, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.025968283414840698, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025315115228295326, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02182777412235737, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020409289747476578, "qparams": { "group_size": 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": 0.0179207231849432, "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": 0.01760067790746689, "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": 0.015281295403838158, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011872473172843456, "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": 0.11978105455636978, "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": 0.11978105455636978, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.32.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.22522717714309692, "qparams": { "group_size": 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": 0.21082071959972382, "qparams": { "group_size": 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": 0.20631317794322968, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1866261214017868, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10635008662939072, "qparams": { "group_size": 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": 0.10112372040748596, "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": 0.11823074519634247, "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": 0.10945440828800201, "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": 0.10791660100221634, "qparams": { "group_size": 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": 0.09522679448127747, "qparams": { "group_size": 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": 0.09004242718219757, "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": 0.06032219156622887, "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": 0.052446164190769196, "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": 0.05103258043527603, "qparams": { "group_size": 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": 0.05070139840245247, "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": 0.030162403360009193, "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": 0.026324450969696045, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026229334995150566, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.024075426161289215, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023864539340138435, "qparams": { "group_size": 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": 0.015907589346170425, "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": 0.016010839492082596, "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": 0.015447713434696198, "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": 0.010744168423116207, "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": 0.11823074519634247, "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": 0.11823074519634247, "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": 0.27842965722084045, "qparams": { "group_size": 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": 0.26073750853538513, "qparams": { "group_size": 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": 0.255304753780365, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.23100101947784424, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.13138447701931, "qparams": { "group_size": 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": 0.12504729628562927, "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": 0.14585289359092712, "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": 0.1350727677345276, "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": 0.133314311504364, "qparams": { "group_size": 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": 0.11772279441356659, "qparams": { "group_size": 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": 0.1112557128071785, "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": 0.07430065423250198, "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": 0.06459391117095947, "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": 0.06290865689516068, "qparams": { "group_size": 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": 0.06251246482133865, "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": 0.03708219528198242, "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": 0.0320652611553669, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03195331245660782, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02923228219151497, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028978103771805763, "qparams": { "group_size": 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": 0.019249066710472107, "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": 0.01890362985432148, "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": 0.018690554425120354, "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": 0.011910264380276203, "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": 0.11772279441356659, "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": 0.11772279441356659, "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": 0.2672635614871979, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24041850864887238, "qparams": { "group_size": 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": 0.2286972999572754, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20545083284378052, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12223910540342331, "qparams": { "group_size": 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": 0.11149773746728897, "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": 0.14622239768505096, "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": 0.13466639816761017, "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": 0.12692317366600037, "qparams": { "group_size": 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": 0.1076633632183075, "qparams": { "group_size": 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": 0.10183274000883102, "qparams": { "group_size": 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": 0.07451047748327255, "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": 0.06448227912187576, "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": 0.05885552614927292, "qparams": { "group_size": 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": 0.057464469224214554, "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": 0.03763241693377495, "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": 0.03067946434020996, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030303651466965675, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027781877666711807, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02690093033015728, "qparams": { "group_size": 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": 0.020411942154169083, "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": 0.01977212354540825, "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": 0.018612347543239594, "qparams": { "group_size": 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": 0.013358589261770248, "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": 0.12223910540342331, "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": 0.12223910540342331, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.14619985222816467, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13633820414543152, "qparams": { "group_size": 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": 0.13245244324207306, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1194988340139389, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06864618510007858, "qparams": { "group_size": 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": 0.06473265588283539, "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": 0.07807900011539459, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07217130064964294, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06973683089017868, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.061302218586206436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05798005312681198, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03966069966554642, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0345073826611042, "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": 0.0328977033495903, "qparams": { "group_size": 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": 0.03251410648226738, "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": 0.019814210012555122, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.016824917867779732, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01667078584432602, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015310825780034065, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015068264678120613, "qparams": { "group_size": 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": 0.010283934883773327, "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": 0.010131621733307838, "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": 0.009754467755556107, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006418304983526468, "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": 0.1194988340139389, "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": 0.1194988340139389, "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": 0.12733972072601318, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11876727640628815, "qparams": { "group_size": 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": 0.1152782216668129, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1039954200387001, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05977553129196167, "qparams": { "group_size": 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": 0.0563163198530674, "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": 0.06808938086032867, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06295749545097351, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.060728106647729874, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05338209867477417, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.050490036606788635, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03458722308278084, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03008781559765339, "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": 0.02865174040198326, "qparams": { "group_size": 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": 0.028300365433096886, "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": 0.017289848998188972, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014662470668554306, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0145178884267807, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013343565165996552, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013124492950737476, "qparams": { "group_size": 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": 0.008984853513538837, "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": 0.00885934941470623, "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": 0.008517157286405563, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0056345099583268166, "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": 0.11876727640628815, "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": 0.11876727640628815, "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": 0.28918200731277466, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2703400254249573, "qparams": { "group_size": 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": 0.26430320739746094, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.2389049232006073, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1362917721271515, "qparams": { "group_size": 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": 0.1294238269329071, "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": 0.1522720456123352, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.14068087935447693, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13838325440883636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.12190979719161987, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.1151033490896225, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07735785096883774, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06719458103179932, "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": 0.06519944220781326, "qparams": { "group_size": 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": 0.06472484767436981, "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": 0.038598544895648956, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03308483958244324, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03293630853295326, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.03006632626056671, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02976243570446968, "qparams": { "group_size": 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": 0.01981264539062977, "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": 0.019310127943754196, "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": 0.01908124051988125, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011783989146351814, "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": 0.12190979719161987, "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": 0.12190979719161987, "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": 0.23866510391235352, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.19756583869457245, "qparams": { "group_size": 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": 0.18119311332702637, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.15592266619205475, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10837727785110474, "qparams": { "group_size": 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": 0.09140502661466599, "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": 0.13359981775283813, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12149044126272202, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.11367923021316528, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08432488143444061, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0794338583946228, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06816292554140091, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05885448306798935, "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": 0.05299503356218338, "qparams": { "group_size": 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": 0.051534056663513184, "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": 0.034479059278964996, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.028913941234350204, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028407391160726547, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.024473264813423157, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023517044261097908, "qparams": { "group_size": 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": 0.019432177767157555, "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": 0.020336829125881195, "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": 0.017692402005195618, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.015529748052358627, "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": 0.12149044126272202, "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": 0.12149044126272202, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "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": 0.22226008772850037, "qparams": { "group_size": 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": 0.2079414427280426, "qparams": { "group_size": 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": 0.20349743962287903, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18394245207309723, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1049576923251152, "qparams": { "group_size": 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": 0.09981272369623184, "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": 0.11665942519903183, "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": 0.1080731451511383, "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": 0.10655294358730316, "qparams": { "group_size": 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": 0.09394299983978271, "qparams": { "group_size": 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": 0.08873973786830902, "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": 0.05951296165585518, "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": 0.051772553473711014, "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": 0.050379399210214615, "qparams": { "group_size": 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": 0.0500565730035305, "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": 0.02975819818675518, "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": 0.025991788133978844, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025900181382894516, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023755256086587906, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023548055440187454, "qparams": { "group_size": 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": 0.01566278375685215, "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": 0.01580328308045864, "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": 0.015204356983304024, "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": 0.010591644793748856, "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": 0.11665942519903183, "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": 0.11665942519903183, "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": 0.2763916850090027, "qparams": { "group_size": 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": 0.2587794065475464, "qparams": { "group_size": 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": 0.2533166706562042, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.22915026545524597, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1305191069841385, "qparams": { "group_size": 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": 0.1241784617304802, "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": 0.14510001242160797, "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": 0.13421730697155, "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": 0.13243375718593597, "qparams": { "group_size": 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": 0.11690053343772888, "qparams": { "group_size": 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": 0.110438771545887, "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": 0.07392141222953796, "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": 0.06420659273862839, "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": 0.06251397728919983, "qparams": { "group_size": 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": 0.0621202327311039, "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": 0.036948636174201965, "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": 0.03193073347210884, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03181964531540871, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.029121985659003258, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.028862137347459793, "qparams": { "group_size": 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": 0.019317423924803734, "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": 0.01892148144543171, "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": 0.01876148022711277, "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": 0.01205119676887989, "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": 0.11690053343772888, "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": 0.11690053343772888, "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": 0.2690129578113556, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.24067263305187225, "qparams": { "group_size": 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": 0.22805410623550415, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20512598752975464, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12265314906835556, "qparams": { "group_size": 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": 0.11119630932807922, "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": 0.14788857102394104, "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": 0.13615670800209045, "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": 0.1277434229850769, "qparams": { "group_size": 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": 0.1077846810221672, "qparams": { "group_size": 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": 0.1022147536277771, "qparams": { "group_size": 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": 0.07513318955898285, "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": 0.06518715620040894, "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": 0.05909130722284317, "qparams": { "group_size": 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": 0.05755116045475006, "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": 0.03797236829996109, "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": 0.03080778941512108, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030393239110708237, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.027859365567564964, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0268926490098238, "qparams": { "group_size": 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": 0.020570339635014534, "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": 0.01996496319770813, "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": 0.01865687593817711, "qparams": { "group_size": 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": 0.013455965556204319, "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": 0.12265314906835556, "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": 0.12265314906835556, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.14921464025974274, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1390499621629715, "qparams": { "group_size": 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": 0.13499948382377625, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12177841365337372, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0700802430510521, "qparams": { "group_size": 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": 0.06602254509925842, "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": 0.07957859337329865, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07373823970556259, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0712120458483696, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06249929964542389, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.059014324098825455, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.040361516177654266, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03522065281867981, "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": 0.0335889495909214, "qparams": { "group_size": 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": 0.0331912636756897, "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": 0.020185938104987144, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.017189742997288704, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.017029641196131706, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015636570751667023, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01538577489554882, "qparams": { "group_size": 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": 0.010488538071513176, "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": 0.010378774255514145, "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": 0.009957514703273773, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.00661660498008132, "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": 0.12177841365337372, "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": 0.12177841365337372, "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": 0.12583576142787933, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11730755120515823, "qparams": { "group_size": 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": 0.11373429000377655, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10251633822917938, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05906134098768234, "qparams": { "group_size": 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": 0.05553540587425232, "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": 0.06758411973714828, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06245005130767822, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.060008205473423004, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05269685760140419, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.049831993877887726, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.034299347549676895, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029854804277420044, "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": 0.028329405933618546, "qparams": { "group_size": 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": 0.02795466221868992, "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": 0.01716332882642746, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014526705257594585, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014372073113918304, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01322386134415865, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01298518292605877, "qparams": { "group_size": 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": 0.008940817788243294, "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": 0.008834470994770527, "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": 0.00844550970941782, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0056627243757247925, "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": 0.11730755120515823, "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": 0.11730755120515823, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.2920745015144348, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.27288076281547546, "qparams": { "group_size": 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": 0.26682376861572266, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.24089083075523376, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.13765032589435577, "qparams": { "group_size": 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": 0.1307055503129959, "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": 0.15384748578071594, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.14204192161560059, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.1397162228822708, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.12298489362001419, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.1161922812461853, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07823515683412552, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06787840276956558, "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": 0.06587231159210205, "qparams": { "group_size": 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": 0.06540042906999588, "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": 0.03903038427233696, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03344644233584404, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03329566866159439, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.03038707748055458, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030079372227191925, "qparams": { "group_size": 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": 0.02004435285925865, "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": 0.01955757476389408, "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": 0.019296299666166306, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01199362426996231, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.1161922812461853, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, "best_option": { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.1161922812461853, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "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": 0.2204686403274536, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18363872170448303, "qparams": { "group_size": 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": 0.16653594374656677, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.13272005319595337, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09889732301235199, "qparams": { "group_size": 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": 0.08359002321958542, "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": 0.12586753070354462, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11585959792137146, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10500643402338028, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0765715092420578, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0713757649064064, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06431441754102707, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05547039210796356, "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": 0.047727301716804504, "qparams": { "group_size": 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": 0.04573405161499977, "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": 0.03216440975666046, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02502288669347763, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.024282265454530716, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02068260684609413, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.019289126619696617, "qparams": { "group_size": 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": 0.017068497836589813, "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": 0.016890164464712143, "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": 0.01455448567867279, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011300390586256981, "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": 0.11585959792137146, "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": 0.11585959792137146, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.34.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.22015328705310822, "qparams": { "group_size": 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": 0.20591136813163757, "qparams": { "group_size": 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": 0.20150312781333923, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18213731050491333, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10400069504976273, "qparams": { "group_size": 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": 0.09885288774967194, "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": 0.115554578602314, "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": 0.10704221576452255, "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": 0.10553359240293503, "qparams": { "group_size": 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": 0.09303726255893707, "qparams": { "group_size": 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": 0.08787233382463455, "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": 0.058934248983860016, "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": 0.05131113529205322, "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": 0.049920160323381424, "qparams": { "group_size": 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": 0.04958723112940788, "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": 0.029494961723685265, "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": 0.025780094787478447, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02568485587835312, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023560967296361923, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02335171401500702, "qparams": { "group_size": 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": 0.015526866540312767, "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": 0.015716886147856712, "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": 0.015070234425365925, "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": 0.010583864524960518, "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": 0.115554578602314, "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": 0.115554578602314, "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": 0.26784318685531616, "qparams": { "group_size": 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": 0.2506856620311737, "qparams": { "group_size": 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": 0.24538299441337585, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.22188712656497955, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12646813690662384, "qparams": { "group_size": 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": 0.12026665359735489, "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": 0.14058224856853485, "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": 0.13006249070167542, "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": 0.12831434607505798, "qparams": { "group_size": 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": 0.11319997161626816, "qparams": { "group_size": 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": 0.1069384440779686, "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": 0.07162083685398102, "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": 0.0622120127081871, "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": 0.06056295707821846, "qparams": { "group_size": 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": 0.060170482844114304, "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": 0.035778723657131195, "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": 0.030898092314600945, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.030787039548158646, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028159843757748604, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027910472825169563, "qparams": { "group_size": 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": 0.018653294071555138, "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": 0.018253767862915993, "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": 0.018111752346158028, "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": 0.011542276479303837, "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": 0.11319997161626816, "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": 0.11319997161626816, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.2661716938018799, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.2367226928472519, "qparams": { "group_size": 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": 0.223912313580513, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.2013011872768402, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12098082154989243, "qparams": { "group_size": 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": 0.10919662564992905, "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": 0.1459568738937378, "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": 0.13458509743213654, "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": 0.1263839602470398, "qparams": { "group_size": 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": 0.10592605173587799, "qparams": { "group_size": 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": 0.10035080462694168, "qparams": { "group_size": 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": 0.07424420863389969, "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": 0.06436290591955185, "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": 0.05822733789682388, "qparams": { "group_size": 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": 0.0567081943154335, "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": 0.03720515966415405, "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": 0.03029671683907509, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02991127409040928, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02730439603328705, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02633078023791313, "qparams": { "group_size": 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": 0.019768686965107918, "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": 0.01956455036997795, "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": 0.017772724851965904, "qparams": { "group_size": 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": 0.013065692037343979, "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": 0.12098082154989243, "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": 0.12098082154989243, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.14398232102394104, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.13423927128314972, "qparams": { "group_size": 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": 0.13035890460014343, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11746101081371307, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06761225312948227, "qparams": { "group_size": 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": 0.06371134519577026, "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": 0.07673574239015579, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.07108988612890244, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0686885342001915, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06028461456298828, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.056892696768045425, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03893664851784706, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.033983081579208374, "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": 0.03241399675607681, "qparams": { "group_size": 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": 0.032036591321229935, "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": 0.01946646347641945, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.016592776402831078, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.016439996659755707, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015091616660356522, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014849630184471607, "qparams": { "group_size": 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": 0.010115830227732658, "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": 0.010026807896792889, "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": 0.009605690836906433, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006403496954590082, "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": 0.11746101081371307, "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": 0.11746101081371307, "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": 0.12170716375112534, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11343692988157272, "qparams": { "group_size": 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": 0.1099933311343193, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09912726283073425, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05707170069217682, "qparams": { "group_size": 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": 0.05374550074338913, "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": 0.06523361057043076, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.060278989374637604, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0580001100897789, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05093168467283249, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04817046970129013, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03311548009514809, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02879953570663929, "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": 0.027367183938622475, "qparams": { "group_size": 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": 0.027026912197470665, "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": 0.01656498946249485, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014033184386789799, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013895129784941673, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012774275615811348, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012555268593132496, "qparams": { "group_size": 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": 0.00864266138523817, "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": 0.008526732213795185, "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": 0.008175482973456383, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005475936457514763, "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": 0.12170716375112534, "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": 0.12170716375112534, "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": 0.2902427017688751, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.27135470509529114, "qparams": { "group_size": 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": 0.2651675343513489, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.23946121335029602, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.13693788647651672, "qparams": { "group_size": 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": 0.13004247844219208, "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": 0.15284109115600586, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.14122694730758667, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13899412751197815, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.12230030447244644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11545172333717346, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07767229527235031, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06747414171695709, "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": 0.06550174206495285, "qparams": { "group_size": 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": 0.0650312751531601, "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": 0.03876513987779617, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.033263761550188065, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0331219844520092, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.030216287821531296, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02991562709212303, "qparams": { "group_size": 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": 0.019921327009797096, "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": 0.019451405853033066, "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": 0.019199352711439133, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011936702765524387, "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": 0.12230030447244644, "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": 0.12230030447244644, "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": 0.2274121344089508, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.18827290832996368, "qparams": { "group_size": 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": 0.17376014590263367, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.1465284377336502, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.10308896750211716, "qparams": { "group_size": 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": 0.08714281767606735, "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": 0.12471625208854675, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.11468721926212311, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.10856347531080246, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.08007597923278809, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.07346683740615845, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.06397257000207901, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.05512146279215813, "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": 0.04987334460020065, "qparams": { "group_size": 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": 0.048571448773145676, "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": 0.032057471573352814, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.026392053812742233, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025992630049586296, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021962886676192284, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021070949733257294, "qparams": { "group_size": 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": 0.017333215102553368, "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": 0.01760166883468628, "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": 0.015674006193876266, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012485479936003685, "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": 0.11468721926212311, "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": 0.11468721926212311, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.35.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.2174573540687561, "qparams": { "group_size": 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": 0.20333951711654663, "qparams": { "group_size": 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": 0.19894367456436157, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.17984575033187866, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.10271351784467697, "qparams": { "group_size": 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": 0.09758777916431427, "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": 0.11409828066825867, "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": 0.10575969517230988, "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": 0.10426422208547592, "qparams": { "group_size": 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": 0.0918298065662384, "qparams": { "group_size": 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": 0.08673278987407684, "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": 0.058101948350667953, "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": 0.05066544562578201, "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": 0.04927938058972359, "qparams": { "group_size": 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": 0.04894713684916496, "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": 0.029042843729257584, "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": 0.025394117459654808, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025295831263065338, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.023182684555649757, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022974630817770958, "qparams": { "group_size": 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": 0.015183573588728905, "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": 0.015393868088722229, "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": 0.014729229733347893, "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": 0.010257821530103683, "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": 0.11409828066825867, "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": 0.11409828066825867, "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": 0.25968000292778015, "qparams": { "group_size": 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": 0.24299459159374237, "qparams": { "group_size": 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": 0.23782654106616974, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.21496005356311798, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12257969379425049, "qparams": { "group_size": 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": 0.11656484752893448, "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": 0.1360863745212555, "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": 0.1261095255613327, "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": 0.1244191974401474, "qparams": { "group_size": 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": 0.10968166589736938, "qparams": { "group_size": 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": 0.1035393625497818, "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": 0.06926912814378738, "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": 0.06030423566699028, "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": 0.05869452655315399, "qparams": { "group_size": 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": 0.05830793455243111, "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": 0.0345681756734848, "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": 0.029880410060286522, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029773728922009468, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02720416709780693, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.026958610862493515, "qparams": { "group_size": 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": 0.017861656844615936, "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": 0.017565567046403885, "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": 0.017324086278676987, "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": 0.010983122512698174, "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": 0.12257969379425049, "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": 0.12257969379425049, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.35.mlp.down_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1715892650462965, "total_bits": 153702656.0, "err": 0.26582765579223633, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.23452816903591156, "qparams": { "group_size": 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": 0.22120477259159088, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.19911149144172668, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.12034080177545547, "qparams": { "group_size": 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": 0.10790876299142838, "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": 0.14566656947135925, "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": 0.13433197140693665, "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": 0.1262228935956955, "qparams": { "group_size": 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": 0.10496708005666733, "qparams": { "group_size": 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": 0.09959550946950912, "qparams": { "group_size": 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": 0.07400427013635635, "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": 0.0641733705997467, "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": 0.05786385387182236, "qparams": { "group_size": 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": 0.056293390691280365, "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": 0.03703294321894646, "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": 0.030029837042093277, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029650336131453514, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.026960015296936035, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02594912052154541, "qparams": { "group_size": 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": 0.019509298726916313, "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": 0.019325103610754013, "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": 0.01744043454527855, "qparams": { "group_size": 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": 0.012730698101222515, "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": 0.12034080177545547, "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": 0.12034080177545547, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.13938356935977936, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12974148988723755, "qparams": { "group_size": 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": 0.12570154666900635, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11318466067314148, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.06532537937164307, "qparams": { "group_size": 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": 0.06143458932638168, "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": 0.07456336915493011, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06907086074352264, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.06638585776090622, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.05815846472978592, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.054940637201070786, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03784427046775818, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03301287442445755, "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": 0.031324081122875214, "qparams": { "group_size": 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": 0.030920151621103287, "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": 0.01892855204641819, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.016057906672358513, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015885325148701668, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.014590037986636162, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014329558238387108, "qparams": { "group_size": 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": 0.009857677854597569, "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": 0.009752870537340641, "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": 0.009319406934082508, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006239517591893673, "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": 0.11318466067314148, "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": 0.11318466067314148, "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": 0.11525781452655792, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10724429041147232, "qparams": { "group_size": 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": 0.10377145558595657, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0933937132358551, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.053972065448760986, "qparams": { "group_size": 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": 0.050651974976062775, "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": 0.0619969479739666, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.0573793388903141, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.054862577468156815, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.048089027404785156, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04545196518301964, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.031450070440769196, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02740221843123436, "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": 0.025881784036755562, "qparams": { "group_size": 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": 0.025510819628834724, "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": 0.015735020861029625, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013277922756969929, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013115654699504375, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012074503116309643, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.011837494559586048, "qparams": { "group_size": 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": 0.008209755644202232, "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": 0.008113620802760124, "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": 0.007723385002464056, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005208847112953663, "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": 0.11525781452655792, "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": 0.11525781452655792, "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": 0.28382357954978943, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.26517608761787415, "qparams": { "group_size": 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": 0.25918933749198914, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.23394262790679932, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.1338529884815216, "qparams": { "group_size": 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": 0.12702859938144684, "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": 0.14974118769168854, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13819004595279694, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13589274883270264, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11953532695770264, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.11278840154409409, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07613645493984222, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06604358553886414, "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": 0.0640721395611763, "qparams": { "group_size": 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": 0.06359991431236267, "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": 0.03799690306186676, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03258237615227699, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.032437678426504135, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.029613204300403595, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.029308591037988663, "qparams": { "group_size": 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": 0.019546445459127426, "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": 0.01915019564330578, "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": 0.018808748573064804, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01186610572040081, "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": 0.11953532695770264, "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": 0.11953532695770264, "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": 0.20613205432891846, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.17224431037902832, "qparams": { "group_size": 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": 0.16013284027576447, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.12575848400592804, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.09404777735471725, "qparams": { "group_size": 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": 0.08092158287763596, "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": 0.11280269175767899, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.10344304889440536, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.0990322083234787, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.07129797339439392, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.06343191862106323, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05800621956586838, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04981240630149841, "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": 0.04554242268204689, "qparams": { "group_size": 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": 0.04448600485920906, "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": 0.029157670214772224, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.024089209735393524, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023801269009709358, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.0197063609957695, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.018968721851706505, "qparams": { "group_size": 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": 0.015819881111383438, "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": 0.015984009951353073, "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": 0.014436276629567146, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011400800198316574, "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": 0.11280269175767899, "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": 0.11280269175767899, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.36.mlp.gate_proj", "numel": 70778880, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1751085069444445, "total_bits": 153951744.0, "err": 0.20769742131233215, "qparams": { "group_size": 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": 0.19428783655166626, "qparams": { "group_size": 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": 0.1900777518749237, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.17173907160758972, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09811633080244064, "qparams": { "group_size": 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": 0.09322074800729752, "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": 0.1091126874089241, "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": 0.10106584429740906, "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": 0.09957651793956757, "qparams": { "group_size": 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": 0.08773372322320938, "qparams": { "group_size": 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": 0.08288532495498657, "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": 0.055576588958501816, "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": 0.0484376884996891, "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": 0.04708696901798248, "qparams": { "group_size": 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": 0.046769317239522934, "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": 0.027794137597084045, "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": 0.024302495643496513, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02421082928776741, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.022204888984560966, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02200520969927311, "qparams": { "group_size": 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": 0.014574501663446426, "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": 0.014804579317569733, "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": 0.014125957153737545, "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": 0.009946856647729874, "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": 0.1091126874089241, "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": 0.1091126874089241, "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": 0.24353773891925812, "qparams": { "group_size": 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": 0.22789081931114197, "qparams": { "group_size": 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": 0.22301232814788818, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.20161500573158264, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1149800643324852, "qparams": { "group_size": 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": 0.10931752622127533, "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": 0.12777136266231537, "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": 0.1183299571275711, "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": 0.11669653654098511, "qparams": { "group_size": 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": 0.10283572971820831, "qparams": { "group_size": 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": 0.09705537557601929, "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": 0.06507010757923126, "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": 0.05659465491771698, "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": 0.05506325513124466, "qparams": { "group_size": 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": 0.05470158904790878, "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": 0.03248744085431099, "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": 0.028070859611034393, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027968529611825943, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.025567354634404182, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025331003591418266, "qparams": { "group_size": 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": 0.01684194803237915, "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": 0.01656711846590042, "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": 0.016326986253261566, "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": 0.010444067418575287, "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": 0.1183299571275711, "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": 0.1183299571275711, "qparams": { "group_size": 32, "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": 0.2546814978122711, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.22275568544864655, "qparams": { "group_size": 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": 0.20963045954704285, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18864509463310242, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1153465285897255, "qparams": { "group_size": 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": 0.10262338072061539, "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": 0.14013591408729553, "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": 0.1283763349056244, "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": 0.12108902633190155, "qparams": { "group_size": 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": 0.09956198930740356, "qparams": { "group_size": 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": 0.0944405272603035, "qparams": { "group_size": 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": 0.0712694376707077, "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": 0.06150686740875244, "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": 0.05564721301198006, "qparams": { "group_size": 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": 0.054190244525671005, "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": 0.035805054008960724, "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": 0.029191115871071815, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02889418974518776, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02609098143875599, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.025167273357510567, "qparams": { "group_size": 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": 0.019225487485527992, "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": 0.019146766513586044, "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": 0.017300700768828392, "qparams": { "group_size": 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": 0.013188730925321579, "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": 0.12108902633190155, "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": 0.12108902633190155, "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": 0.12969091534614563, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.12066706269979477, "qparams": { "group_size": 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": 0.11683894693851471, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10512685775756836, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0607682541012764, "qparams": { "group_size": 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": 0.05701111629605293, "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": 0.06968032568693161, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06438439339399338, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.061745937913656235, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.054062020033597946, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05109955742955208, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.035387441515922546, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.030798980966210365, "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": 0.029153870418667793, "qparams": { "group_size": 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": 0.028761126101017, "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": 0.017723482102155685, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.015026265755295753, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.014859610237181187, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013668157160282135, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013418608345091343, "qparams": { "group_size": 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": 0.009314488619565964, "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": 0.009258894249796867, "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": 0.008796031586825848, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.006087905261665583, "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": 0.12066706269979477, "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": 0.12066706269979477, "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": 0.11032146960496902, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.10269845277070999, "qparams": { "group_size": 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": 0.09919977933168411, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08923875540494919, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.05165531486272812, "qparams": { "group_size": 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": 0.048378173261880875, "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": 0.05967393517494202, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.055151306092739105, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05248389020562172, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04601559042930603, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.0435035340487957, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03028925694525242, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.02637445740401745, "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": 0.024798410013318062, "qparams": { "group_size": 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": 0.024417953565716743, "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": 0.01517176441848278, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012773911468684673, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012606375850737095, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011626026593148708, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01138631347566843, "qparams": { "group_size": 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": 0.00797125045210123, "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": 0.007895665243268013, "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": 0.007464882917702198, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.0051582627929747105, "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": 0.11032146960496902, "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": 0.11032146960496902, "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": 0.26872339844703674, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.25083988904953003, "qparams": { "group_size": 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": 0.24499057233333588, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.22090253233909607, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12664934992790222, "qparams": { "group_size": 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": 0.12009306252002716, "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": 0.14187704026699066, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13098274171352386, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12863123416900635, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11297481507062912, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.1065969243645668, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07212590426206589, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06256480515003204, "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": 0.06062348186969757, "qparams": { "group_size": 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": 0.06014528498053551, "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": 0.035996753722429276, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030866436660289764, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.03071404993534088, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02802731841802597, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027727186679840088, "qparams": { "group_size": 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": 0.018548332154750824, "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": 0.018199577927589417, "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": 0.01784161850810051, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011348976753652096, "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": 0.11297481507062912, "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": 0.11297481507062912, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.16694200038909912, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.14276133477687836, "qparams": { "group_size": 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": 0.13286913931369781, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.11099416017532349, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07582846283912659, "qparams": { "group_size": 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": 0.06571120023727417, "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": 0.09217981994152069, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08443226665258408, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.07886605709791183, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.06062855198979378, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05597401037812233, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.04734240472316742, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.04149774834513664, "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": 0.03770917281508446, "qparams": { "group_size": 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": 0.03677014634013176, "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": 0.024147799238562584, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.021605588495731354, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021277694031596184, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019015338271856308, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.018449394032359123, "qparams": { "group_size": 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": 0.01400366984307766, "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": 0.016203340142965317, "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": 0.012899458408355713, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013409778475761414, "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": 0.11099416017532349, "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": 0.11099416017532349, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "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": 0.19605734944343567, "qparams": { "group_size": 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": 0.18330778181552887, "qparams": { "group_size": 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": 0.1792423129081726, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1619364321231842, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.09257401525974274, "qparams": { "group_size": 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": 0.08795648068189621, "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": 0.1031072661280632, "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": 0.09539872407913208, "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": 0.09398139268159866, "qparams": { "group_size": 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": 0.08275499194860458, "qparams": { "group_size": 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": 0.07819445431232452, "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": 0.052637699991464615, "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": 0.04577907174825668, "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": 0.04448595643043518, "qparams": { "group_size": 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": 0.04417679086327553, "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": 0.02635442093014717, "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": 0.02305467799305916, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022961607202887535, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021086012944579124, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020888596773147583, "qparams": { "group_size": 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": 0.013968248851597309, "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": 0.014192109927535057, "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": 0.013537028804421425, "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": 0.009712069295346737, "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": 0.1031072661280632, "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": 0.1031072661280632, "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": 0.22258251905441284, "qparams": { "group_size": 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": 0.20814970135688782, "qparams": { "group_size": 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": 0.20362895727157593, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18404406309127808, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.1050492525100708, "qparams": { "group_size": 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": 0.09982810914516449, "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": 0.11692710220813751, "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": 0.10816742479801178, "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": 0.1066262498497963, "qparams": { "group_size": 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": 0.09393063932657242, "qparams": { "group_size": 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": 0.08867451548576355, "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": 0.05959989130496979, "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": 0.05178651958703995, "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": 0.05035551264882088, "qparams": { "group_size": 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": 0.050014231353998184, "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": 0.029786689206957817, "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": 0.025779815390706062, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02568097971379757, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02349168434739113, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.023280322551727295, "qparams": { "group_size": 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": 0.015575463883578777, "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": 0.015375158749520779, "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": 0.015097484923899174, "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": 0.009913926012814045, "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": 0.11692710220813751, "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": 0.11692710220813751, "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": 0.24513480067253113, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.21338526904582977, "qparams": { "group_size": 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": 0.20110183954238892, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.18062549829483032, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11123722046613693, "qparams": { "group_size": 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": 0.09882699698209763, "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": 0.13423708081245422, "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": 0.12277377396821976, "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": 0.11663204431533813, "qparams": { "group_size": 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": 0.09523884952068329, "qparams": { "group_size": 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": 0.09034141153097153, "qparams": { "group_size": 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": 0.06851464509963989, "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": 0.05905191972851753, "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": 0.053859736770391464, "qparams": { "group_size": 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": 0.05257866904139519, "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": 0.03458163142204285, "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": 0.02858068235218525, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02833825908601284, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02549581602215767, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.024681005626916885, "qparams": { "group_size": 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": 0.01895829848945141, "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": 0.019073281437158585, "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": 0.017319004982709885, "qparams": { "group_size": 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": 0.01368545088917017, "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": 0.12277377396821976, "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": 0.12277377396821976, "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": 0.12575691938400269, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11711204797029495, "qparams": { "group_size": 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": 0.11336375027894974, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10209104418754578, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.058919236063957214, "qparams": { "group_size": 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": 0.05534779652953148, "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": 0.06773312389850616, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06249431520700455, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.059861332178115845, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.0525214858353138, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04976922646164894, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03440041467547417, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.029880907386541367, "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": 0.0282891858369112, "qparams": { "group_size": 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": 0.02790832705795765, "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": 0.017241906374692917, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.014572362415492535, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01440976932644844, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.013277247548103333, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013032772578299046, "qparams": { "group_size": 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": 0.00908111035823822, "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": 0.008971069939434528, "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": 0.008584942668676376, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005878380034118891, "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": 0.11711204797029495, "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": 0.11711204797029495, "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": 0.10592672973871231, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.09863942861557007, "qparams": { "group_size": 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": 0.09523458778858185, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.08575672656297684, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04959443211555481, "qparams": { "group_size": 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": 0.046445656567811966, "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": 0.0573463961482048, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.05302906781435013, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05039528012275696, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.04420005530118942, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04183242842555046, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.029090646654367447, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.025350570678710938, "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": 0.023811787366867065, "qparams": { "group_size": 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": 0.023430753499269485, "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": 0.01457536406815052, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.012265476398169994, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012095890007913113, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011176069267094135, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010938157327473164, "qparams": { "group_size": 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": 0.007675379514694214, "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": 0.007589009124785662, "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": 0.007189438212662935, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004958919249475002, "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": 0.10592672973871231, "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": 0.10592672973871231, "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": 0.27489155530929565, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.2568875551223755, "qparams": { "group_size": 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": 0.251135915517807, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.2266356498003006, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12962737679481506, "qparams": { "group_size": 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": 0.12305543571710587, "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": 0.14485648274421692, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.13366524875164032, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.13157916069030762, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.11569396406412125, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10930255800485611, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.0736754909157753, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.06388479471206665, "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": 0.06204144284129143, "qparams": { "group_size": 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": 0.061599455773830414, "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": 0.036747995764017105, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.03158078342676163, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0314503088593483, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.028693247586488724, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.0284140482544899, "qparams": { "group_size": 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": 0.018933333456516266, "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": 0.018578071147203445, "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": 0.018251288682222366, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.011568160727620125, "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": 0.11569396406412125, "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": 0.11569396406412125, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "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": 0.17358598113059998, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1425696164369583, "qparams": { "group_size": 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": 0.1299763172864914, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.10849544405937195, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.07875135540962219, "qparams": { "group_size": 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": 0.06562946736812592, "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": 0.09757000207901001, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.08872558176517487, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.08295787125825882, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.060463301837444305, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.05535265430808067, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.05045546218752861, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.043493568897247314, "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": 0.03903160244226456, "qparams": { "group_size": 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": 0.03795228525996208, "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": 0.02577964775264263, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.02208489179611206, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021730996668338776, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01889496110379696, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.018192946910858154, "qparams": { "group_size": 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": 0.015000293031334877, "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": 0.016348080709576607, "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": 0.013716517947614193, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.013240644708275795, "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": 0.10849544405937195, "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": 0.10849544405937195, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "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": 0.16366639733314514, "qparams": { "group_size": 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": 0.15291287004947662, "qparams": { "group_size": 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": 0.14955921471118927, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.13517671823501587, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.0775737389922142, "qparams": { "group_size": 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": 0.07369481772184372, "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": 0.08645111322402954, "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": 0.07992912083864212, "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": 0.0787353366613388, "qparams": { "group_size": 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": 0.0694580152630806, "qparams": { "group_size": 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": 0.06573270261287689, "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": 0.04457159340381622, "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": 0.039110977202653885, "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": 0.03804488107562065, "qparams": { "group_size": 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": 0.03778858855366707, "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": 0.022517958655953407, "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": 0.020978635177016258, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020914899185299873, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.019493190571665764, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.019341690465807915, "qparams": { "group_size": 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": 0.012754992581903934, "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": 0.014565723016858101, "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": 0.01242807973176241, "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": 0.011748461984097958, "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": 0.08645111322402954, "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": 0.08645111322402954, "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": 0.177946999669075, "qparams": { "group_size": 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": 0.1663806140422821, "qparams": { "group_size": 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": 0.16272017359733582, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.14718525111675262, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08436935395002365, "qparams": { "group_size": 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": 0.08025017380714417, "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": 0.094077929854393, "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": 0.08694300800561905, "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": 0.08563680201768875, "qparams": { "group_size": 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": 0.07559669017791748, "qparams": { "group_size": 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": 0.07161245495080948, "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": 0.04857151582837105, "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": 0.04249535873532295, "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": 0.04136613756418228, "qparams": { "group_size": 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": 0.04110071063041687, "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": 0.02458546869456768, "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": 0.022795476019382477, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022724226117134094, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.021182037889957428, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.021021662279963493, "qparams": { "group_size": 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": 0.014027156867086887, "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": 0.01579108089208603, "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": 0.013682753778994083, "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": 0.01271712128072977, "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": 0.094077929854393, "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": 0.094077929854393, "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": 0.18932510912418365, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.1633959710597992, "qparams": { "group_size": 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": 0.15385444462299347, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.13766999542713165, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.08568393439054489, "qparams": { "group_size": 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": 0.0759151428937912, "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": 0.10447558015584946, "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": 0.0944473072886467, "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": 0.08973497897386551, "qparams": { "group_size": 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": 0.07306233048439026, "qparams": { "group_size": 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": 0.06942754983901978, "qparams": { "group_size": 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": 0.05354219675064087, "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": 0.04603245481848717, "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": 0.042020607739686966, "qparams": { "group_size": 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": 0.0410287119448185, "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": 0.027448473498225212, "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": 0.023150354623794556, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.022923888638615608, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02084430120885372, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.020197072997689247, "qparams": { "group_size": 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": 0.01591903530061245, "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": 0.016424227505922318, "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": 0.014570011757314205, "qparams": { "group_size": 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": 0.012783135287463665, "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": 0.10447558015584946, "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": 0.10447558015584946, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.39.self_attn.q_proj", "numel": 26214400, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.17529296875, "total_bits": 57024000.0, "err": 0.12073332071304321, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.11259078234434128, "qparams": { "group_size": 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": 0.1088581383228302, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.09806300699710846, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.0564984567463398, "qparams": { "group_size": 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": 0.05302005261182785, "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": 0.06546485424041748, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06021280586719513, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.05739125981926918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.050430648028850555, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04787285253405571, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.033255014568567276, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0287984199821949, "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": 0.0271115954965353, "qparams": { "group_size": 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": 0.026704205200076103, "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": 0.016650648787617683, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.013959109783172607, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.013780619949102402, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.012728343717753887, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012470693327486515, "qparams": { "group_size": 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": 0.008758362382650375, "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": 0.008619583211839199, "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": 0.008225638419389725, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.005630075465887785, "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": 0.12073332071304321, "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": 0.12073332071304321, "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": 0.09604545682668686, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.0894780382514, "qparams": { "group_size": 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": 0.08629707992076874, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.07771331071853638, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.04486457258462906, "qparams": { "group_size": 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": 0.0419553779065609, "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": 0.05215317755937576, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.04818126559257507, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.045586250722408295, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.040035780519247055, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.037902966141700745, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.026450837031006813, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.023052366450428963, "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": 0.021543089300394058, "qparams": { "group_size": 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": 0.021178312599658966, "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": 0.013271618634462357, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.011131987906992435, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010965364053845406, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01015542447566986, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.00992478895932436, "qparams": { "group_size": 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": 0.0070197428576648235, "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": 0.006959838327020407, "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": 0.006553161423653364, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.004602620378136635, "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": 0.09604545682668686, "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": 0.09604545682668686, "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": 0.2573868930339813, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.24073591828346252, "qparams": { "group_size": 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": 0.2353125512599945, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.21250811219215393, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.12155541032552719, "qparams": { "group_size": 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": 0.11534576117992401, "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": 0.13735373318195343, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.12555146217346191, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.12335807085037231, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.10871214419603348, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.10315645486116409, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.07017394155263901, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.0602944977581501, "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": 0.058368317782878876, "qparams": { "group_size": 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": 0.057874299585819244, "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": 0.035499099642038345, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.030019376426935196, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.02983933314681053, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.02739972434937954, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.027081701904535294, "qparams": { "group_size": 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": 0.01858583092689514, "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": 0.01815752312541008, "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": 0.01780220866203308, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.01183826383203268, "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": 0.12155541032552719, "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": 0.12155541032552719, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "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": 0.1345479041337967, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.37529296875, "total_bits": 62266880.00000001, "err": 0.1017657145857811, "qparams": { "group_size": 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": 0.08944996446371078, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.72529296875, "total_bits": 71441920.0, "err": 0.0789375975728035, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.22529296875, "total_bits": 84549120.0, "err": 0.059260044246912, "qparams": { "group_size": 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": 0.04397788643836975, "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": 0.07681921869516373, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.12529296875, "total_bits": 81927680.0, "err": 0.06720145046710968, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.17529296875, "total_bits": 83238400.0, "err": 0.063850037753582, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.52529296875, "total_bits": 92413440.0, "err": 0.042910050600767136, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.662646484375, "total_bits": 96014080.0, "err": 0.04175703972578049, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.0313232421875, "total_bits": 105678720.0, "err": 0.03962118923664093, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.12529296875, "total_bits": 108142080.0, "err": 0.03374388441443443, "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": 0.030288876965641975, "qparams": { "group_size": 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": 0.02942686900496483, "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": 0.021173648536205292, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.22529296875, "total_bits": 136977920.0, "err": 0.01828940398991108, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.018104609102010727, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.015715545043349266, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015224484726786613, "qparams": { "group_size": 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": 0.013396044261753559, "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": 0.014506056904792786, "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": 0.012499748729169369, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.12529296875, "total_bits": 212999679.99999997, "err": 0.012561498209834099, "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": 0.1017657145857811, "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": 0.1017657145857811, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "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": 0.14525160193443298, "qparams": { "group_size": 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": 0.13563309609889984, "qparams": { "group_size": 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": 0.13260334730148315, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.11986484378576279, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.06827597320079803, "qparams": { "group_size": 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": 0.06483586132526398, "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": 0.07612640410661697, "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": 0.07040343433618546, "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": 0.0693131685256958, "qparams": { "group_size": 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": 0.06105145439505577, "qparams": { "group_size": 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": 0.057703763246536255, "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": 0.03879257291555405, "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": 0.03373440355062485, "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": 0.032776858657598495, "qparams": { "group_size": 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": 0.03255050629377365, "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": 0.019402042031288147, "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": 0.016958758234977722, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.016893619671463966, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01550045982003212, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.015355001203715801, "qparams": { "group_size": 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": 0.010220160707831383, "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": 0.010396012105047703, "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": 0.009898735210299492, "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": 0.007058984134346247, "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": 0.11986484378576279, "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": 0.11986484378576279, "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": 0.1415121704339981, "qparams": { "group_size": 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": 0.13222752511501312, "qparams": { "group_size": 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": 0.1292589157819748, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.116808220744133, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.06695321202278137, "qparams": { "group_size": 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": 0.06357678771018982, "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": 0.07477620244026184, "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": 0.06901926547288895, "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": 0.06796786189079285, "qparams": { "group_size": 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": 0.05992067977786064, "qparams": { "group_size": 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": 0.05678008124232292, "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": 0.038547977805137634, "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": 0.033526863902807236, "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": 0.03259837254881859, "qparams": { "group_size": 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": 0.0323796384036541, "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": 0.019435636699199677, "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": 0.017618361860513687, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01755758747458458, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.01629417948424816, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.01616268791258335, "qparams": { "group_size": 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": 0.010880137793719769, "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": 0.011824728921055794, "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": 0.010598092339932919, "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": 0.009178119711577892, "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": 0.116808220744133, "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": 0.116808220744133, "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": 0.0914192721247673, "qparams": { "group_size": 32, "bits": [ 3, 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": 0.07902049273252487, "qparams": { "group_size": 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": 0.07358942925930023, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.06575258076190948, "qparams": { "group_size": 32, "bits": [ 4, 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": 0.04154668748378754, "qparams": { "group_size": 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": 0.03640948608517647, "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": 0.05325362831354141, "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": 0.04708551615476608, "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": 0.04354766011238098, "qparams": { "group_size": 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": 0.0355297289788723, "qparams": { "group_size": 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": 0.03409543260931969, "qparams": { "group_size": 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": 0.02733750455081463, "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": 0.023323632776737213, "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": 0.02086270973086357, "qparams": { "group_size": 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": 0.020241299644112587, "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": 0.014213496819138527, "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": 0.012218886986374855, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.012062504887580872, "qparams": { "group_size": 32, "bits": [ 8, 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": 0.011205591261386871, "qparams": { "group_size": 32, "bits": [ 6, 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": 0.010862786322832108, "qparams": { "group_size": 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": 0.008511756546795368, "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": 0.009427467361092567, "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": 0.007740021217614412, "qparams": { "group_size": 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": 0.007948262616991997, "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": 0.0914192721247673, "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": 0.0914192721247673, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } } ], "last_module_idx": 82, "base_perplexity": 7.725421353299703 }