{ "measurement": [ { "key": "model.layers.0.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.041570764034986496, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.04118853434920311, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.020015548914670944, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.01995515078306198, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.019923603162169456, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.0069623468443751335, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.04223167151212692, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.0409812293946743, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.019956428557634354, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.01989135704934597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.020025011152029037, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.020365087315440178, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.019880330190062523, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.010730935260653496, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.006941793020814657, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.01086756307631731, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.006932680960744619, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.005331153050065041, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.00693088723346591, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.00532795675098896, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.007006875239312649, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.006930591072887182, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.005377314984798431, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.005327058956027031, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.041570764034986496, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.041570764034986496, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.03556209057569504, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.03529679402709007, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.016632085666060448, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.01659441739320755, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.016572380438447, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.007012472487986088, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.037331368774175644, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.035128358751535416, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.0165906623005867, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.01654730923473835, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.016661236062645912, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.017517514526844025, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.016539182513952255, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.009583888575434685, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.00700162211433053, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.009677917696535587, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.006996089592576027, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.005995146930217743, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.0069944774731993675, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.005994334351271391, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.00704596983268857, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.0069951945915818214, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.006017942912876606, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005994067527353764, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.03556209057569504, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.03556209057569504, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.11731315404176712, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.10889872908592224, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.06156102195382118, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.05816817656159401, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.056303564459085464, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.022433128207921982, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1288335621356964, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.10461172461509705, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.058343350887298584, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.05385701358318329, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.06531314551830292, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07366456836462021, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.05282743275165558, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.026743251830339432, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.01635601744055748, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.04029613733291626, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.013502526096999645, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.007653325796127319, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.012988673523068428, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.006271986290812492, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.01859922520816326, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.01277866493910551, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.006107751280069351, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005437290295958519, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11731315404176712, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11731315404176712, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.00740050571039319, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.005030300002545118, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.0033509323839098215, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.0030969223007559776, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.0025273426435887814, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.0015859929844737053, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.004934271797537804, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.004377747420221567, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.0032977410592138767, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.002325469860807061, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.002401270205155015, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.002416827715933323, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.0021758447401225567, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.0014554503140971065, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.0012208421248942614, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.0013267082395032048, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.0010782802710309625, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.0010217807721346617, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.0010578997898846865, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.0009568971581757069, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.0008886757423169911, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.0010445729130879045, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.0007167597650550306, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.0009308187873102725, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.00740050571039319, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.00740050571039319, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.16596150398254395, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.1528911143541336, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.1456964910030365, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.12850670516490936, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.07598716020584106, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.0711294412612915, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.09160294383764267, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.08423157036304474, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.07797516882419586, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.0680430456995964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.0617738775908947, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.045884326100349426, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.040059007704257965, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.03650382533669472, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.03562313690781593, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.02299966849386692, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.019053582102060318, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.018674161285161972, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.017488986253738403, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.01650521717965603, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01227407157421112, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.012343604117631912, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01111401803791523, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.008389616385102272, "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": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.12850670516490936, "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": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.12850670516490936, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2038041204214096, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.18776948750019073, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.18105120956897736, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.15977199375629425, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.09379048645496368, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.08872849494218826, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1089748814702034, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.10064470767974854, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.09615089744329453, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.08390270173549652, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.07596354186534882, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.055163491517305374, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.047932498157024384, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.04486804082989693, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.04412557929754257, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.027522552758455276, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.02302495762705803, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.022738205268979073, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.021025696769356728, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.01998502016067505, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.014281199313700199, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01414951030164957, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.013178467750549316, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.009059945121407509, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1089748814702034, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1089748814702034, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.0.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.09065176546573639, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.06772752106189728, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.05741254985332489, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.04920953884720802, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.03888586536049843, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.028825342655181885, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.05066731572151184, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.046499159187078476, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.04208691790699959, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.02792886644601822, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.027187451720237732, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.025751357898116112, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.02252179943025112, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.019169701263308525, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.018333276733756065, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.01309238187968731, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.010867048054933548, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.010603372007608414, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.0090031074360013, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.008531334809958935, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.007401252631098032, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.008212250657379627, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.006412217393517494, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.006551219616085291, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.09065176546573639, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.09065176546573639, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.06298412382602692, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.06145394966006279, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.029296405613422394, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.02868640050292015, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.028060318902134895, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.015202600508928299, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.0691053569316864, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.05999905243515968, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.028375746682286263, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.027740782126784325, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.028909733518958092, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.030276237055659294, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.027501892298460007, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.017691951245069504, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.01458718255162239, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.017464470118284225, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.014387937262654305, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.013286394998431206, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.014358065091073513, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.013241413980722427, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.013711104169487953, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.014341761358082294, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.01225993037223816, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.013210703618824482, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.06298412382602692, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.06298412382602692, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.05437528342008591, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.05360895395278931, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.02098742127418518, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.020684964954853058, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.020313922315835953, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.010574529878795147, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.06568074226379395, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.05258120596408844, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.020451344549655914, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.02017929218709469, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.020960189402103424, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.02506696805357933, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.02005067840218544, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.012687133625149727, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.010297033935785294, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.012288623489439487, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.010212631896138191, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.0095746461302042, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.010199771262705326, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.009554922580718994, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.009815432131290436, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.01019207201898098, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.009164630435407162, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.009541560895740986, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.05437528342008591, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.05437528342008591, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.16595135629177094, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.1377214640378952, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.11760835349559784, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.09927234053611755, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.07587883621454239, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.05871716886758804, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.11856791377067566, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.098964162170887, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.08156267553567886, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.06446313112974167, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.0650186613202095, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.06519642472267151, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.05231758579611778, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.03959137201309204, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.03346744552254677, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.037461694329977036, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.02150113880634308, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.01684506982564926, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.01947496086359024, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.013264366425573826, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.023489464074373245, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.017528444528579712, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.011297554709017277, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.00837064441293478, "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": 2.380859375, "total_bits": 1248256.0, "err": 0.1377214640378952, "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": 2.380859375, "total_bits": 1248256.0, "err": 0.1377214640378952, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.1.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.1765785664319992, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.12914350628852844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.09874111413955688, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.08292673528194427, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.07020646333694458, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.04973195493221283, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.11358669400215149, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.10266495496034622, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.08341080695390701, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.05671785771846771, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.05589383468031883, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.05811459943652153, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.04979436844587326, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.035509783774614334, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.03128349781036377, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.02935185097157955, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.021026000380516052, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.01972988434135914, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.018857106566429138, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.016205478459596634, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.01657465100288391, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.017830628901720047, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.012238417752087116, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.014020662754774094, "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": 2.375732421875, "total_bits": 9964544.0, "err": 0.12914350628852844, "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": 2.375732421875, "total_bits": 9964544.0, "err": 0.12914350628852844, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.1840672641992569, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.1703743040561676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.16543903946876526, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.14715924859046936, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.08390951156616211, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.08011806756258011, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.0953640565276146, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.08816680312156677, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.085792176425457, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.07569455355405807, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.06903241574764252, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.04851190745830536, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.04210055246949196, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.04020799323916435, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.03975340723991394, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.024260036647319794, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.02103397250175476, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.0208954606205225, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.01943349465727806, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.01871406100690365, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.013005692511796951, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.013345766812562943, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.012400387786328793, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.009383206255733967, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.0953640565276146, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.0953640565276146, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.22848038375377655, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.21311679482460022, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.20783881843090057, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.18584983050823212, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1051076129078865, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10098299384117126, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.11852148175239563, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.10957736521959305, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.10731364786624908, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.0953732430934906, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.08723098784685135, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.060015663504600525, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05213005840778351, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05015299841761589, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.04968530684709549, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.029928049072623253, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.025546623393893242, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.025407040491700172, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.023569565266370773, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.022689471021294594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.015478183515369892, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.015179278329014778, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.014800714328885078, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.009485675022006035, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.11852148175239563, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.11852148175239563, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.1.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.21642421185970306, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.19684815406799316, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.18932293355464935, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.16995878517627716, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.09934763610363007, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.09143128991127014, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.11555177718400955, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.10621879994869232, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.10164915025234222, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.08695298433303833, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.0831245705485344, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.05889555439352989, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.05094585940241814, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.047873690724372864, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.04712696745991707, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.029581304639577866, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.025328343734145164, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.025067487731575966, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.02295197919011116, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.022658661007881165, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.016053665429353714, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.016595328226685524, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.015079839155077934, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.0120177511125803, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.11555177718400955, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.11555177718400955, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.05040634050965309, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.046561747789382935, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.030123967677354813, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.027047330513596535, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.022709986194968224, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.013712561689317226, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.04604288190603256, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.040838852524757385, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.02352660708129406, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.02117692120373249, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.021745363250374794, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.022230764850974083, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.019396817311644554, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.011597177013754845, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.00894437450915575, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.0113567179068923, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.0070912535302340984, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.005859510973095894, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.006851132959127426, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.005463020410388708, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.006553425453603268, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.006585291121155024, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.004508228972554207, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.0049749636091291904, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.05040634050965309, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.05040634050965309, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.037034790962934494, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.03544031083583832, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.018329089507460594, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.01703214831650257, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.01547298114746809, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.008367895148694515, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.03765004128217697, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.03304382786154747, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.015801168978214264, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.014904113486409187, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.015199776738882065, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.016395194455981255, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.01431133784353733, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.008696231059730053, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.00670696422457695, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.008417167700827122, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.006149223539978266, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.005400205962359905, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.006080567836761475, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.005298834294080734, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.005831483285874128, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.006012363359332085, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.004807785153388977, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005184700712561607, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.037034790962934494, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.037034790962934494, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.20410694181919098, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.1771375983953476, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.16052986681461334, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.13757270574569702, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.09355555474758148, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.08018551766872406, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1389986127614975, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.11323250830173492, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.09860577434301376, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.07983506470918655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.07817310839891434, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.0747138112783432, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.05637736991047859, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.045642610639333725, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.042252231389284134, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.039495740085840225, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.02344820275902748, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.021813631057739258, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02070968598127365, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.017770810052752495, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.02143959328532219, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.016165928915143013, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.01306509505957365, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.009316656738519669, "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": 2.755859375, "total_bits": 1444864.0, "err": 0.13757270574569702, "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": 2.755859375, "total_bits": 1444864.0, "err": 0.13757270574569702, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.21257241070270538, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.16685496270656586, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.14492744207382202, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.11752398312091827, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.09264932572841644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.07609357684850693, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.12641285359859467, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.11383210122585297, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.10126110166311264, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.07020365446805954, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.06657131761312485, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.06503191590309143, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.05594344437122345, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.04625840112566948, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.0437377467751503, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03308985382318497, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.02672087773680687, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.025912495329976082, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.023013580590486526, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.021065842360258102, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.018713127821683884, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.02089839056134224, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.015741271898150444, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.016875751316547394, "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": 2.750732421875, "total_bits": 11537408.0, "err": 0.11752398312091827, "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": 2.750732421875, "total_bits": 11537408.0, "err": 0.11752398312091827, "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": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.23980769515037537, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2227703332901001, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.21739040315151215, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.19396176934242249, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11199958622455597, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10748554021120071, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12646514177322388, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.11629114300012589, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1142401397228241, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10131553560495377, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09256071597337723, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06489886343479156, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.056988999247550964, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.055053479969501495, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05457735061645508, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.033021293580532074, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03065863996744156, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03055541031062603, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.028735505416989326, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.027911759912967682, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.018862752243876457, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.021637041121721268, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.018246745690703392, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.017667043954133987, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12646514177322388, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12646514177322388, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.26200348138809204, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.24438944458961487, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.23867055773735046, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.2132161259651184, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.12218699604272842, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11743909120559692, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13729937374591827, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12697938084602356, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12468638271093369, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.11065720021724701, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.1009717509150505, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07012179493904114, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06081574410200119, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.058681972324848175, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05817333236336708, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.035118333995342255, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.030451923608779907, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.030311184003949165, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.028132738545536995, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02716008387506008, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.018709251657128334, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.018863879144191742, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.018007081001996994, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.01292490679770708, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13729937374591827, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13729937374591827, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.2.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.0811636820435524, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.07754826545715332, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.048065491020679474, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.044645894318819046, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.03536876663565636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.02297982946038246, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.08016172051429749, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.06847242265939713, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.03582208976149559, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.03352905437350273, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.032475825399160385, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.029974447563290596, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.030249731615185738, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.02334226854145527, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.01623929664492607, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.020520586520433426, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.013830810785293579, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.006182831712067127, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.013599189929664135, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.00567457964643836, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.014010006561875343, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.013191068544983864, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.004778298549354076, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.004043652676045895, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.0811636820435524, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.0811636820435524, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.056699272245168686, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.05071895197033882, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.04452615976333618, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.038874268531799316, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.02586449310183525, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.022042492404580116, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.03665344417095184, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.03368741646409035, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.02696257270872593, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.022743158042430878, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.021296825259923935, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.018389085307717323, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.016099240630865097, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.012595642358064651, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.011631355620920658, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.009230908937752247, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.00673759076744318, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.0062677436508238316, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.006143125705420971, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.005395200103521347, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.004942918196320534, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.004951868671923876, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.003870765445753932, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.0033471263013780117, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.056699272245168686, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.056699272245168686, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.04979880526661873, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.044171929359436035, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.03701423481106758, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.03235276788473129, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.022281518206000328, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.018145335838198662, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.03413717448711395, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.03125777468085289, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.02343425340950489, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.019731441512703896, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.018919840455055237, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.016795862466096878, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.014828762039542198, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.010914178565144539, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.009760189801454544, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.008465577848255634, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.005847867112606764, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.005303129553794861, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.005356408189982176, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.0045683979988098145, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.004504833370447159, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.00446854904294014, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.003382052294909954, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.002979376818984747, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.04979880526661873, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.04979880526661873, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.22903166711330414, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.20576441287994385, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.19407127797603607, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.16921062767505646, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.10542163997888565, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.09660879522562027, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1340201497077942, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.11869499832391739, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.10944627225399017, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.09233451634645462, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.08547542244195938, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.06940148025751114, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.0570051483809948, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05059606209397316, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.049039945006370544, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.034790847450494766, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.02572607435286045, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.02506413497030735, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.022979220375418663, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.021182719618082047, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.01767122559249401, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.016000064089894295, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.014571502804756165, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.009323803707957268, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1340201497077942, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1340201497077942, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.20700062811374664, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.17523285746574402, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.1630786955356598, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.13499772548675537, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.0915670394897461, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.0824877917766571, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.11374663561582565, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.10335884988307953, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.0980009213089943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.07496843487024307, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.06791695207357407, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.05833318084478378, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.04999658837914467, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.044739168137311935, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.04335915669798851, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.02957807295024395, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.024181708693504333, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.02383800968527794, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.0210361797362566, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.019531114026904106, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.016406839713454247, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.016836758702993393, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.014756145887076855, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.012589965015649796, "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": 2.750732421875, "total_bits": 11537408.0, "err": 0.13499772548675537, "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": 2.750732421875, "total_bits": 11537408.0, "err": 0.13499772548675537, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2608211934566498, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2424899935722351, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.23633527755737305, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.21071098744869232, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.12191540002822876, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11686501652002335, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13746599853038788, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1271456778049469, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12458183616399765, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10981260985136032, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09991173446178436, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0700397938489914, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06080881506204605, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05844496190547943, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05787951499223709, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03503137826919556, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.029951902106404305, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.029781140387058258, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02747999131679535, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.026406172662973404, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.018398402258753777, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01803630031645298, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01760394684970379, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011591262184083462, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13746599853038788, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13746599853038788, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.28670525550842285, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2669057548046112, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2604014277458191, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23217521607875824, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13425984978675842, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.12871244549751282, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.15102329850196838, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13984215259552002, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13710299134254456, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12098447978496552, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11003467440605164, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07678362727165222, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.0667991191148758, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06426874548196793, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06368456780910492, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03835413232445717, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03274712339043617, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.0325724221765995, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.03001367300748825, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02883341535925865, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.019832946360111237, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01941635087132454, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.018984289839863777, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.01209736242890358, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13710299134254456, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13710299134254456, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.3.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.1301819384098053, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.11474991589784622, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.1055154800415039, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.09469745308160782, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.0628204420208931, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.052493106573820114, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.08445117622613907, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.0727435052394867, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.06485288590192795, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.05464533716440201, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.05321861803531647, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.0475722961127758, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.03987306356430054, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.03478580340743065, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.03063141368329525, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.02773725613951683, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.020506424829363823, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.018414825201034546, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.01927001215517521, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.017010919749736786, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.01733998954296112, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.017295973375439644, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.013875707983970642, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.013341930694878101, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.1301819384098053, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.1301819384098053, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.08368712663650513, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.07605119049549103, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.070949986577034, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.062030594795942307, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.03864022344350815, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.03526441752910614, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.049438219517469406, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.044860560446977615, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.039942946285009384, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.03400710970163345, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.03134537115693092, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.0250377394258976, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.021415460854768753, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.01864747330546379, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.01795324683189392, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.012537935748696327, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.009704156778752804, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.009388117119669914, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.008772457018494606, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.008096903562545776, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.006558524444699287, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.006409910507500172, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.005618310067802668, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.004195710178464651, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.08368712663650513, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.08368712663650513, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.08498695492744446, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.07604949921369553, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.06876526772975922, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.05997295305132866, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.038276441395282745, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.03379512578248978, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.05140506103634834, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.0473104827105999, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.039805032312870026, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.03357284516096115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.03113727644085884, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.025738539174199104, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.022320007905364037, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.018455728888511658, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.017439186573028564, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.012869455851614475, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.009558765217661858, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.009026688523590565, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.008621628396213055, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.007675913628190756, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.0066503072157502174, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.006480011157691479, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.005304032936692238, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.003932623192667961, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.08498695492744446, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.08498695492744446, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.2452956736087799, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.22358404099941254, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.2135077863931656, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.18766027688980103, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.11329341679811478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.10581900924444199, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1372070014476776, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12421850115060806, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.11684553325176239, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.100147545337677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.09157449007034302, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07002165913581848, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.05926906317472458, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.054257676005363464, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.05308075249195099, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.03511698544025421, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.027601664885878563, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.027065914124250412, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.024879205971956253, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.023299075663089752, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.017893921583890915, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.01678786240518093, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.015776164829730988, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.009893777780234814, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1372070014476776, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1372070014476776, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.2517118752002716, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.21174439787864685, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.19343873858451843, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.16066882014274597, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11174583435058594, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.09796614199876785, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.14441868662834167, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1317318081855774, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12046363949775696, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.0923876017332077, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.08252110332250595, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07406283169984818, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.06350917369127274, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05443207174539566, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.05213072523474693, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.037130653858184814, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.029286909848451614, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.02851499617099762, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.02555926889181137, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.02327742427587509, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.02001849189400673, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.020573802292346954, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.01708693616092205, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.014805507846176624, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1317318081855774, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1317318081855774, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.26631245017051697, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.24753712117671967, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.24143251776695251, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.21515163779258728, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.12489193677902222, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.1196586936712265, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.14060844480991364, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13022691011428833, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1276700496673584, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.11243738234043121, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.10211516171693802, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07162979245185852, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.062304653227329254, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05988205224275589, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.059305377304553986, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.0357917919754982, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.030695008113980293, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.030533209443092346, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02813483588397503, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.027045072987675667, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.018657442182302475, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01848609745502472, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01785680465400219, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011899081990122795, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13022691011428833, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13022691011428833, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.29407963156700134, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2736629545688629, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2669481933116913, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23810522258281708, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13800004124641418, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.13222740590572357, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1552664041519165, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14371740818023682, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.14099879562854767, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12430055439472198, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11302817612886429, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0789242535829544, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06869681924581528, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06608512997627258, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06545758247375488, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.039426468312740326, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.033609237521886826, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03344474360346794, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.030780484899878502, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02955538220703602, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020372288301587105, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.019845956936478615, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01949162967503071, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012229252606630325, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13800004124641418, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13800004124641418, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.4.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.27856969833374023, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.24763453006744385, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2355431318283081, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.20902875065803528, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1297142207622528, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.11640061438083649, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.1522773951292038, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.13993118703365326, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13350452482700348, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.10953180491924286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.10386332869529724, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.07757779955863953, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.06725791096687317, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06262809038162231, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06147259846329689, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.03896545246243477, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.032835979014635086, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.0324208028614521, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.028846336528658867, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.028384190052747726, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.020869603380560875, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.021179258823394775, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.01935088261961937, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.01476652454584837, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13350452482700348, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13350452482700348, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.0749819278717041, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.06846864521503448, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.06380893290042877, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.05620035156607628, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.03473605215549469, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.03163949400186539, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.04589849337935448, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.04058348760008812, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.03583128750324249, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.030842212960124016, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.028508933261036873, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.023378340527415276, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.01938282512128353, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.016749627888202667, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.016078338027000427, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.011680196039378643, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.008683686144649982, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.008366074413061142, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.007895349524915218, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.00725571159273386, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.0060717081651091576, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.005714144557714462, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.004968556575477123, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.0036462238058447838, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.0749819278717041, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.0749819278717041, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.07466772198677063, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.06750568002462387, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.0616849847137928, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.05382157117128372, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.0340222492814064, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.03030151128768921, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.04526088759303093, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.041276220232248306, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.03521772846579552, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.03004492074251175, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.02789156511425972, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.022717604413628578, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.019684849306941032, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.016359271481633186, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.015527643263339996, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.011412698775529861, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.008419635705649853, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.008005037903785706, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.007614660542458296, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.006859084591269493, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.005826900713145733, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.005591947585344315, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.004691889509558678, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.0033701416105031967, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.07466772198677063, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.07466772198677063, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.25076058506965637, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.22851106524467468, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.21839049458503723, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.19224520027637482, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.11591805517673492, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.10821977257728577, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.14131121337413788, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12728725373744965, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.1193944662809372, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.1028706431388855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.09459125250577927, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07235812395811081, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.060886602848768234, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05567421391606331, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.054354049265384674, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.03623795136809349, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.028075989335775375, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.027558529749512672, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.025356678292155266, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02372344769537449, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.018427878618240356, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.01695483922958374, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.01607125997543335, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.009742734022438526, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12728725373744965, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12728725373744965, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.25223758816719055, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.22033533453941345, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.2083035707473755, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.1741672158241272, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11550028622150421, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.10579993575811386, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13795694708824158, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.126420259475708, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12059233337640762, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.09649510681629181, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.08528328686952591, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07085015624761581, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.061387479305267334, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05640549957752228, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.055161867290735245, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03566722199320793, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.030570490285754204, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.030217833817005157, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.027013638988137245, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.025347769260406494, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.019514769315719604, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.02112080529332161, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.017912333831191063, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.01606648787856102, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13795694708824158, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13795694708824158, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2731372117996216, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2538653612136841, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2474033385515213, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.22029925882816315, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.128337562084198, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.12282621115446091, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1448795348405838, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1341233104467392, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13128352165222168, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.11549954116344452, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.10486233979463577, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0738907977938652, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06426335871219635, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.061621226370334625, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06100443750619888, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03695861995220184, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.031756941229104996, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03157978504896164, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.029131369665265083, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.028012176975607872, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01934916339814663, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.019396429881453514, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.018474237993359566, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012820358388125896, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1341233104467392, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1341233104467392, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.30260127782821655, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2815946638584137, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.27463287115097046, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.24461521208286285, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1422470659017563, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.13625505566596985, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.16021665930747986, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14832854270935059, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1453951895236969, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.1280500590801239, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11631535738706589, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0815504863858223, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.07091342657804489, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06814326345920563, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06747069954872131, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.04070967808365822, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03466806933283806, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03447533771395683, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.031718797981739044, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.030449550598859787, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020964190363883972, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.020502867177128792, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.019999125972390175, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012640343979001045, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.1280500590801239, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.1280500590801239, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.5.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.2850662171840668, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.25290340185165405, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2406924068927765, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.21430924534797668, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1326705664396286, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.11887279897928238, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.1549965888261795, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.14309090375900269, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13676993548870087, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11228736490011215, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.10673901438713074, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.07934024930000305, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.06888394057750702, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06405092030763626, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06286896020174026, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.03985573723912239, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03354984149336815, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.033129774034023285, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.029511380940675735, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.029013188555836678, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.02131878212094307, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.02160213328897953, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.019751887768507004, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.015007463283836842, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13676993548870087, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13676993548870087, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.07588335126638412, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.0683390349149704, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.06218956038355827, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.054624270647764206, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.03480027988553047, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.03084353357553482, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.04700176417827606, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.04265785589814186, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.03626706823706627, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.030787073075771332, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.028682466596364975, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.02379578724503517, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.020390689373016357, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.01688263565301895, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.015962567180395126, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.011911392211914062, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.008938496932387352, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.008510355837643147, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.0081506771966815, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.007379735354334116, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.0062772659584879875, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.006253786385059357, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.00507122278213501, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.004227296449244022, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.07588335126638412, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.07588335126638412, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.07761182636022568, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.0689045861363411, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.06087758019566536, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.05310545116662979, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.03496302664279938, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.02991650253534317, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.04904824495315552, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.04512954503297806, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.0365106575191021, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.030723147094249725, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.028982460498809814, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.024814067408442497, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.0215273667126894, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.016914093866944313, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.015681929886341095, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.012426386587321758, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.008948957547545433, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.00835594441741705, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.00811952818185091, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.0071546598337590694, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.006585448049008846, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.00643506646156311, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.005128753371536732, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.004212637897580862, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.07761182636022568, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.07761182636022568, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.23179367184638977, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.21087795495986938, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.20180745422840118, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.17725466191768646, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.10746791213750839, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.09999313950538635, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.13248667120933533, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.11765068769454956, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.11070271581411362, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.09492721408605576, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.08713807910680771, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.06793487817049026, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.056447286158800125, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.051529448479413986, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.050376564264297485, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.03383953124284744, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.026255829259753227, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.025718776509165764, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.023621389642357826, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.022131899371743202, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.017300166189670563, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.016067788004875183, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.0149787412956357, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.009567482396960258, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.13248667120933533, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.13248667120933533, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.21796315908432007, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.19160224497318268, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.17897194623947144, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.15275295078754425, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.09935536980628967, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.0895109623670578, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.12488473206758499, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.11308033019304276, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.10388050228357315, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.0849578008055687, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.07613424211740494, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.06420519948005676, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.05512762814760208, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.04885095730423927, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.047286733984947205, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03256792575120926, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.02687298133969307, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.026279650628566742, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.024203216657042503, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.0226067453622818, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.018148938193917274, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.019303252920508385, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.016134830191731453, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.01489709410816431, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.12488473206758499, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.12488473206758499, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2574913501739502, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.23945464193820953, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.23317916691303253, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.20790426433086395, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.12093769758939743, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11574967205524445, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13653172552585602, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12635235488414764, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12365177273750305, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10882049053907394, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09892116487026215, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06966417282819748, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06052389740943909, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.058046068996191025, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.057468824088573456, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.0348205603659153, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.029857678338885307, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.02968605048954487, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02738838456571102, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02630937658250332, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01818155124783516, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.0181704331189394, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.017335709184408188, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011881985701620579, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13653172552585602, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13653172552585602, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2954244315624237, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2749473452568054, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2681077718734741, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23904289305210114, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13884863257408142, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.13306677341461182, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1568976193666458, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14481770992279053, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.14194795489311218, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12508027255535126, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11388538777828217, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07983684539794922, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06927447021007538, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06654175370931625, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06590399146080017, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03985463082790375, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03385690599679947, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.033668942749500275, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.030979519709944725, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.029743507504463196, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.02057330124080181, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.020026380196213722, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.019621778279542923, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012351845391094685, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13884863257408142, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13884863257408142, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.6.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.28332141041755676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.2512686848640442, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.23929184675216675, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.21233049035072327, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13213641941547394, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.11858655512332916, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.1544075310230255, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.14176853001117706, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13612939417362213, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11126093566417694, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.1056690439581871, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.07901225239038467, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.06845393031835556, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06397704035043716, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06290348619222641, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.03983668237924576, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03387151658535004, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.033503878861665726, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.029817869886755943, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.029381657019257545, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.021594636142253876, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.022227557376027107, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.020117662847042084, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.01602892577648163, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13612939417362213, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13612939417362213, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.09767651557922363, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.08952072262763977, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.08324147760868073, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.07356878370046616, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.04518755525350571, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.04116164520382881, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.05845551937818527, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.05321511626243591, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.04665036499500275, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.0404660664498806, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.037489261478185654, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.029561607167124748, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.02546571008861065, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.02189861796796322, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.021005336195230484, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.014819479547441006, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.011561274528503418, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.011124001815915108, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.010627610608935356, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.009793292731046677, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.00788344070315361, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.007884749211370945, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.006689765490591526, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.005376838147640228, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.09767651557922363, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.09767651557922363, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.0852801650762558, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.07777345925569534, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.07152947038412094, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.06309903413057327, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.039114486426115036, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.03503793850541115, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.05094432830810547, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.047326117753982544, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.040319427847862244, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.03491048887372017, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.03233761712908745, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.02564031071960926, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.022594153881072998, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.018854083493351936, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.017914677038788795, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.012805642560124397, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.009844777174293995, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.009352576918900013, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.008994542062282562, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.008139107376337051, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.006751024164259434, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.0066770827397704124, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.005608638748526573, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.004260777961462736, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.0852801650762558, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.0852801650762558, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.2678647041320801, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.24641337990760803, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.23811915516853333, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.2105158120393753, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.1245269924402237, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.11766763776540756, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1484566330909729, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13265526294708252, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.12758302688598633, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.1110374927520752, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.1020154282450676, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07653847336769104, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.06351794302463531, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.0596066489815712, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.058565761893987656, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.03837519511580467, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.030168453231453896, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.02980169467628002, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.027360262349247932, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02596280351281166, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.01955912448465824, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.017945675179362297, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.017358794808387756, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010489912703633308, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13265526294708252, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13265526294708252, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.2115550935268402, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.18923231959342957, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.18144504725933075, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.15319575369358063, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.09704829752445221, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.09075426310300827, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.11404739320278168, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.10387234389781952, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.10053962469100952, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.08272665739059448, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.07261018455028534, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.058563828468322754, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.05008251219987869, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.046996813267469406, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.04624465852975845, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.029422473162412643, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.02488424815237522, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.024684324860572815, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.02210705541074276, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.020927486941218376, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.01586519181728363, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.016353163868188858, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.014856064692139626, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.01187924761325121, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.11404739320278168, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.11404739320278168, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.238688126206398, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2220456451177597, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.21585708856582642, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.1924685537815094, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11232162266969681, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10728666931390762, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1281754970550537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.11806732416152954, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11484614759683609, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10113127529621124, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09219536185264587, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0655747652053833, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05679274722933769, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.054178427904844284, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.053531207144260406, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03282776102423668, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.02831963263452053, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.02811601385474205, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02608364075422287, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.025061873719096184, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017452728003263474, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017962070181965828, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.016520699486136436, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.0125627052038908, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1281754970550537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1281754970550537, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2904992997646332, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2706291675567627, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.26347360014915466, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23532406985759735, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13674922287464142, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.13075989484786987, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.156214639544487, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1436205506324768, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13971532881259918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12337540835142136, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11253612488508224, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07952898740768433, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06872644275426865, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06558559089899063, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.0648377537727356, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03973450884222984, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03347327932715416, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03322099149227142, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.030688920989632607, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.029407544061541557, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020562075078487396, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.02003488875925541, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.019411806017160416, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.01250444259494543, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13674922287464142, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13674922287464142, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.7.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.09381140768527985, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.0871167778968811, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.05739664286375046, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.05210431292653084, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.039071228355169296, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.025922950357198715, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.08132807165384293, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.07429586350917816, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.03991807997226715, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.035705696791410446, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.03462487459182739, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.03187277913093567, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.030502120032906532, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.01766926981508732, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.014865221455693245, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.015858106315135956, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.009532645344734192, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.007721689064055681, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.008831222541630268, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.006824679207056761, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.008351825177669525, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.007775889243930578, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.005193566903471947, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.004393851384520531, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.09381140768527985, "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": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.09381140768527985, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.09052726626396179, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.08118884265422821, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.07219980657100677, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.06388480216264725, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.04154202342033386, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.03575320541858673, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.05966620147228241, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.05316069349646568, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.04341953247785568, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.036927539855241776, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.03512250632047653, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.030364612117409706, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.025581080466508865, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.020185351371765137, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.01871558092534542, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.015238028019666672, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.010660696774721146, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.009961622767150402, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.00975104235112667, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.00862274132668972, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.007909285835921764, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.007653009612113237, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.005961894989013672, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.005014684516936541, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.09052726626396179, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.09052726626396179, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.08618659526109695, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.07645285874605179, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.06497354060411453, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.057343609631061554, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.039104897528886795, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.031589239835739136, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.05966521427035332, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.05369258299469948, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.04113186523318291, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.03472432866692543, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.033396780490875244, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.02999911829829216, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.026049135252833366, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.018829133361577988, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.016875697299838066, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.015064824372529984, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.009837938472628593, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.008821310475468636, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.008960631676018238, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.007451476063579321, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.007703504059463739, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.0072784642688930035, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.00511252461001277, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.004282768350094557, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.08618659526109695, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.08618659526109695, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.24003270268440247, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.21736523509025574, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.20553846657276154, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.18043525516986847, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.11127910017967224, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.10231699049472809, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.14503614604473114, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12504231929779053, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.11531667411327362, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.09823638200759888, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.09070046246051788, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07469311356544495, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.060191985219717026, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.053566571325063705, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.05190504342317581, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.037292297929525375, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.0273257065564394, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.026607085019350052, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02458914741873741, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.022816944867372513, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.01899607479572296, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.017179515212774277, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.015507691539824009, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010332247242331505, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12504231929779053, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12504231929779053, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.24265284836292267, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.2116924524307251, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.1958678662776947, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.1689150482416153, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11071609705686569, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.09866274893283844, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.14228412508964539, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12844166159629822, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.11567538231611252, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.0939989760518074, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.0862671360373497, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07307098060846329, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.062242843210697174, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05418340489268303, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.05215338617563248, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.036927834153175354, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.029518743976950645, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.028703678399324417, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.02634170837700367, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.024445049464702606, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.02031540311872959, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.02099713869392872, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.01774582266807556, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.015787778422236443, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12844166159629822, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12844166159629822, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.25879254937171936, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.24093014001846313, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.23491883277893066, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.20953279733657837, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.12169689685106277, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11659867316484451, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13756665587425232, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1269901841878891, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12441760301589966, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10968334972858429, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09974641352891922, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07023452967405319, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06084693223237991, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05845468118786812, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.057880401611328125, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03509695082902908, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03008659929037094, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.029925473034381866, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.027612382546067238, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.026556165888905525, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01844719797372818, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01829945668578148, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.017623424530029297, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.01200958900153637, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13756665587425232, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13756665587425232, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.30322563648223877, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.28257790207862854, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2758012115955353, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.24624662101268768, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.143072709441185, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.1371188759803772, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.16144880652427673, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14898419380187988, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.14621329307556152, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12898606061935425, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11773128062486649, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.08259820938110352, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.07149234414100647, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06880935281515121, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06814990192651749, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.041328974068164825, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03549891337752342, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.035324614495038986, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.03264657035470009, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.031413301825523376, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.021852020174264908, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.021727075800299644, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.02092299982905388, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.01447279378771782, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12898606061935425, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12898606061935425, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.8.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.2898636758327484, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.2568100690841675, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.24306973814964294, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.2163274586200714, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13510268926620483, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12032055109739304, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16127151250839233, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.14775411784648895, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.1393018364906311, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11428200453519821, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.10899712890386581, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08284564316272736, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07144884765148163, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.065579354763031, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.0641157254576683, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.041746124625205994, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03482036665081978, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.034244924783706665, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.03078801929950714, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.03011077083647251, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.02257591113448143, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.023206911981105804, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.020541826263070107, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.016693711280822754, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.1393018364906311, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.1393018364906311, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.1011006087064743, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.09290925413370132, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.08758781850337982, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.07756437361240387, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.047105491161346436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.04347285255789757, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.05994024500250816, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.05368208885192871, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.04843423515558243, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.04204808920621872, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.0391148179769516, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.030372893437743187, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.025693519040942192, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.0227402001619339, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.02199653908610344, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.01519735250622034, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.011819664388895035, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.011470798403024673, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.010798781178891659, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.010062867775559425, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.007897214964032173, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.007702710572630167, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.006784217432141304, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.005030984990298748, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.1011006087064743, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.1011006087064743, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.08856245130300522, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.08088147640228271, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.07461651414632797, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.06566444784402847, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.04051835089921951, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.0365474596619606, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.05273318290710449, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.04856238514184952, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.041847795248031616, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.03614366054534912, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.03350960463285446, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.02664756216108799, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.023033658042550087, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.019525326788425446, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.01861455664038658, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.013274328783154488, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.010020342655479908, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.009553692303597927, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.009103894233703613, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.008264020085334778, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.0068278927356004715, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.006563762668520212, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.005633347202092409, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.003949633799493313, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.08856245130300522, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.08856245130300522, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.25437477231025696, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.23382508754730225, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.22586996853351593, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.1996867060661316, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.11840201169252396, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.11189307272434235, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.1414409577846527, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12650103867053986, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.12144791334867477, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.1056053638458252, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.09714685380458832, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07276968657970428, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.06051292642951012, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05667309835553169, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.05575335770845413, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.03651238605380058, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.028605276718735695, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.028236733749508858, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02591298706829548, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02457413822412491, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.018546167761087418, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.016839299350976944, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.01639186404645443, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.009694875217974186, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12650103867053986, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12650103867053986, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.26539841294288635, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.23547488451004028, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.22437329590320587, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.1953413188457489, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.12115943431854248, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.11184483021497726, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.14301402866840363, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1318238526582718, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12600596249103546, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10446099191904068, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.09403835237026215, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07291915267705917, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.06300275027751923, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05817243456840515, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.05698558688163757, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.036391016095876694, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.029932402074337006, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.029539896175265312, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.02654465287923813, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.024974029511213303, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.018945034593343735, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.018685046583414078, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.017341582104563713, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.012026488780975342, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1318238526582718, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1318238526582718, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.24634750187397003, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.22921347618103027, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.22308433055877686, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.1991657316684723, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11583593487739563, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11079397797584534, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13124531507492065, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1212543174624443, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11840517073869705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10433849692344666, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09505297988653183, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06687752157449722, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.058075349777936935, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05559077486395836, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05500498414039612, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03342127799987793, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.028529707342386246, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.02835303544998169, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.026175836101174355, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.025120776146650314, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017437996342778206, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017266733571887016, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.016601648181676865, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011155381798744202, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13124531507492065, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13124531507492065, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2972550392150879, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2768709063529968, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2701012194156647, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.241005077958107, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13985946774482727, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.13406772911548615, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1580970585346222, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14601220190525055, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1429385393857956, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12612779438495636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11490412801504135, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.08048838376998901, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06986234337091446, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06706972420215607, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06640773266553879, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.04017159342765808, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.034113191068172455, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.033914171159267426, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.03124857135117054, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.029984813183546066, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020712824538350105, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.020159680396318436, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.019729753956198692, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012386626563966274, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12612779438495636, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12612779438495636, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.9.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.29864025115966797, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.26302433013916016, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.25012144446372986, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.22100912034511566, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1390528827905655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12440676242113113, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16260382533073425, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.14928653836250305, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.14341187477111816, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11614597588777542, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11001672595739365, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08333124220371246, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.0719759538769722, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06736286729574203, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06621377915143967, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.04190893471240997, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.0355038084089756, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.03512841463088989, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.031002547591924667, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.03056388348340988, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.022750038653612137, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.023105675354599953, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.021274177357554436, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.016454296186566353, "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": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1390528827905655, "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": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1390528827905655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.09758958965539932, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.08999931812286377, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.08526851236820221, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.0756177306175232, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.04555618017911911, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.042347948998212814, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.05724379047751427, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.051275357604026794, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.04677356779575348, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.04081762954592705, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.03759654983878136, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.02900102734565735, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.024533748626708984, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.021981393918395042, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.02135176584124565, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.01448323205113411, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.01139409001916647, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.011110388673841953, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.010440408252179623, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.009775103069841862, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.007533177267760038, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.007356696296483278, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.006609863135963678, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.004821718670427799, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.09758958965539932, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.09758958965539932, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.09706295281648636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.08880757540464401, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.08270557969808578, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.07272934168577194, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.04454902186989784, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.040636222809553146, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.05736164376139641, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.05214826017618179, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.04596112668514252, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.03977952525019646, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.0368826724588871, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.028886154294013977, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.024931158870458603, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.021481294184923172, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.020584920421242714, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.014429249800741673, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.011019227094948292, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.010588380508124828, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.01002961490303278, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.0091811902821064, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.007442784029990435, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.007141063921153545, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.006253869738429785, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.004333482589572668, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.09706295281648636, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.09706295281648636, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.26684051752090454, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.24594472348690033, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.23751316964626312, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.21035391092300415, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.12444914877414703, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.11762610822916031, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.14964795112609863, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13324259221553802, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.12764276564121246, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.11119983345270157, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.10338447242975235, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07724684476852417, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.0638175904750824, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05969031900167465, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.05868653580546379, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.0388418585062027, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.03018372319638729, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.02982785925269127, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.0273768100887537, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02594398520886898, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.019698861986398697, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.017904356122016907, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.017359673976898193, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010390966199338436, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13324259221553802, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13324259221553802, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.26327645778656006, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.23677262663841248, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.2265738993883133, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.1948324292898178, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.12089837342500687, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.11281201243400574, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.14255791902542114, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.13092589378356934, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12514597177505493, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10480976849794388, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.0933704674243927, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07308358699083328, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.06299040466547012, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05840713530778885, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.057276125997304916, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03666321188211441, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.030706383287906647, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.03032223880290985, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.02752218022942543, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.025958018377423286, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.01953631453216076, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.019947558641433716, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.017998022958636284, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.014014962129294872, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.13092589378356934, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.13092589378356934, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2507267892360687, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.23350109159946442, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.22747185826301575, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.20304109156131744, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11794193834066391, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11287485808134079, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13349506258964539, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12338416278362274, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12054917216300964, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10634732246398926, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09690167009830475, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06804895401000977, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.059105079621076584, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05660735443234444, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05602327734231949, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03401758149266243, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.02901974879205227, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.028839463368058205, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02663717046380043, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.025576651096343994, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01773262396454811, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017537729814648628, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01687251590192318, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011286848224699497, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13349506258964539, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13349506258964539, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2972685992717743, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.27702030539512634, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2702063322067261, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.24128025770187378, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1399332880973816, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.1340980976819992, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.15805305540561676, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1459244191646576, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1429615169763565, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12615637481212616, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11496143043041229, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.08048011362552643, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06981483101844788, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06707663834095001, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06639940291643143, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.040209997445344925, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03416439890861511, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.033972203731536865, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.03130349516868591, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.030065249651670456, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.02075071446597576, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.02023717761039734, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01977398991584778, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012555349618196487, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12615637481212616, "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": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12615637481212616, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.10.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.3006926476955414, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.26603975892066956, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2531793415546417, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.2240149825811386, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.14066176116466522, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12579166889190674, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16425633430480957, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.1512385606765747, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.14489635825157166, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.1177794337272644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11158274859189987, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08420270681381226, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07289645075798035, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.0680493637919426, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06686577200889587, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.04239886999130249, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03582143411040306, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.0354212187230587, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.03134267032146454, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.030862698331475258, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.023061729967594147, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.023260772228240967, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.021500157192349434, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.016455069184303284, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.1177794337272644, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.1177794337272644, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.11834590882062912, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.10785271972417831, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.10169347375631332, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.08974665403366089, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.05486734211444855, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.050504833459854126, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.06912975758314133, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.06218051537871361, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.0565241202712059, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.048620980232954025, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.045042525976896286, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.03520950675010681, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.029788894578814507, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.026536330580711365, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.025706380605697632, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.01763547956943512, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.013869562186300755, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.01350577361881733, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.01262224093079567, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.011782472021877766, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.009204242378473282, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.009134534746408463, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.007957832887768745, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.006164694670587778, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.11834590882062912, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.11834590882062912, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.1085422933101654, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.09851568192243576, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.09192406386137009, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.08064626902341843, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.049818333238363266, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.04527541249990463, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.06335411965847015, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.05805114284157753, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.05153196305036545, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.044186223298311234, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.0407155305147171, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.0321614034473896, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.027871500700712204, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.023992324247956276, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.023004235699772835, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.016056504100561142, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.012334377504885197, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.011888651177287102, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.011177558451890945, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.010250982828438282, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.008326967246830463, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.008046519011259079, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.0069777523167431355, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.004964878316968679, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.1085422933101654, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.1085422933101654, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.25161516666412354, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.23224608600139618, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.22482845187187195, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.1987851858139038, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.11729196459054947, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.11128602921962738, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.13937626779079437, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.12504905462265015, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.12014181911945343, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.1052200198173523, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.0965985655784607, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.0714995265007019, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.05983160436153412, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05629522353410721, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.05539887398481369, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.03580281883478165, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.02848181128501892, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.028140299022197723, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02593042142689228, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.024664508178830147, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.018229372799396515, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.01689552702009678, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.016375282779335976, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010018776170909405, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.13937626779079437, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.13937626779079437, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.25425153970718384, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.22683154046535492, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.21518872678279877, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.18618479371070862, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11572195589542389, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.10680487006902695, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.14061999320983887, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12870174646377563, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12078197300434113, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10054674744606018, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.091184601187706, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07183554023504257, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.06175714358687401, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05576861649751663, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.0542990081012249, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03593183308839798, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.029172319918870926, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.02864398993551731, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.026163360103964806, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.02447163127362728, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.018990639597177505, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.019022878259420395, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.017015589401125908, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.012924491427838802, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12870174646377563, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12870174646377563, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2397802472114563, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.22299188375473022, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2169523984193802, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.1936158984899521, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1129385232925415, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10789421945810318, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1282266229391098, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.11853092163801193, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11544457077980042, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10170982778072357, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.0926351323723793, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06546125560998917, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05682368203997612, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05426035821437836, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05365750193595886, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.0327133908867836, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.02795657142996788, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.027757227420806885, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.025680847465991974, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.024641813710331917, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017152095213532448, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017121806740760803, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.016277996823191643, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011273439973592758, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1282266229391098, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1282266229391098, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.29572659730911255, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2754621207714081, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.26845842599868774, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23965708911418915, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1393377035856247, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.13346175849437714, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1576627641916275, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14558357000350952, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.14232677221298218, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12555626034736633, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11437462270259857, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.08037248253822327, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06964028626680374, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06680840253829956, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06613828986883163, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.04012288153171539, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.034015756100416183, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03382037952542305, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.031146660447120667, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02990415133535862, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020704880356788635, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.02017984166741371, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.0196855328977108, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012473211623728275, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1393377035856247, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1393377035856247, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.11.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.3043338656425476, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.27038469910621643, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.25780028104782104, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.22767674922943115, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.14255814254283905, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12802107632160187, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16591018438339233, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.15260915458202362, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.14669175446033478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11951657384634018, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11305057257413864, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.0848369300365448, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07350321859121323, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.068882055580616, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06776754558086395, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.04257407784461975, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03606200963258743, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.0356864370405674, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.03150574117898941, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.0310899019241333, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.022700710222125053, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.02312254160642624, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.021182363852858543, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.01613803394138813, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11951657384634018, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11951657384634018, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.12243794649839401, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.11279291659593582, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.10662876814603806, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.09439574182033539, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.05713501200079918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.052880533039569855, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.07214732468128204, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.06463964283466339, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.05860692262649536, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.05106212943792343, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.04725213721394539, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.03683624416589737, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.030944889411330223, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.027514131739735603, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.026668038219213486, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.018391482532024384, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.014270978979766369, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.013861876912415028, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.013047594577074051, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.012170819565653801, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.009503471665084362, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.009248651564121246, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.008078771643340588, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.006022394634783268, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.12243794649839401, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.12243794649839401, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.09585502743721008, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.08834807574748993, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.08167039602994919, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.07196225225925446, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.04420186206698418, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.03995480015873909, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.05671050027012825, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.05296090617775917, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.045424461364746094, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.03960302472114563, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.036555975675582886, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.02865052968263626, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.025240186601877213, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.021314991638064384, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.020308561623096466, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.014328853227198124, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.011063884012401104, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.010547413490712643, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.010128330439329147, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.009220204316079617, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.00755954347550869, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.007390188984572887, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.00636376952752471, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.0046876841224730015, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.09585502743721008, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.09585502743721008, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.2872082591056824, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.26557105779647827, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.257939875125885, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.22843575477600098, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.1345106065273285, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.12792502343654633, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.15810002386569977, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.1417655646800995, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.13760504126548767, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.12016398459672928, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.11047650873661041, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.08165261149406433, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.06778255850076675, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.06432536989450455, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.06349164247512817, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.04102066904306412, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.03241792693734169, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.03212705999612808, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02940947562456131, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.027995074167847633, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.020773259922862053, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.018914448097348213, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.01862192526459694, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010862911120057106, "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": 3.193359375, "total_bits": 1674240.0, "err": 0.13760504126548767, "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": 3.193359375, "total_bits": 1674240.0, "err": 0.13760504126548767, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.12.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.2682080864906311, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.2414274662733078, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.23178799450397491, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.19992117583751678, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.12304285913705826, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.11485280096530914, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.1443442404270172, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1318381130695343, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12728910148143768, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10674840956926346, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.0948815569281578, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07382948696613312, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.06321778893470764, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05925067514181137, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.058265093713998795, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.036933138966560364, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.03092975728213787, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.030640151351690292, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.027673592790961266, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.02620864287018776, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.01966877281665802, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.01972198858857155, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.018350860103964806, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.013620540499687195, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1318381130695343, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1318381130695343, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.22652164101600647, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2107466161251068, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2046576291322708, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.18256118893623352, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.10669432580471039, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10177446901798248, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12202434241771698, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1126394271850586, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1091262474656105, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.09608025848865509, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.087649866938591, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0623900331556797, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05415962263941765, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05145510286092758, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05079973116517067, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03124924935400486, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.026854606345295906, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.02664201520383358, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02473052591085434, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.023742130026221275, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.016614185646176338, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01698155701160431, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01571664772927761, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011778311803936958, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12202434241771698, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12202434241771698, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2910096049308777, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.27084362506866455, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2637859880924225, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23551155626773834, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1370605230331421, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.13107654452323914, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1560257077217102, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14366254210472107, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.14012661576271057, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12354107201099396, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11256265640258789, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07954899221658707, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06880601495504379, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06577502936124802, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06505324691534042, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.039701271802186966, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03355032578110695, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.033324241638183594, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.030733412131667137, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02946464903652668, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020499909296631813, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.020044436678290367, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.019397545605897903, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012543505057692528, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1370605230331421, "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": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1370605230331421, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.12.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.29968222975730896, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.2655331492424011, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2528046667575836, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.22235038876533508, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.14018774032592773, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12570008635520935, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16451269388198853, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.15060274302959442, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.14430761337280273, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.1172000840306282, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11068803817033768, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.0840817391872406, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07267310470342636, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06788105517625809, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06673838943243027, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.04227408766746521, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.035798199474811554, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.03540241718292236, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.03129146620631218, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.03083917312324047, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.022783363237977028, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.023320361971855164, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.021236997097730637, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.016608482226729393, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.1172000840306282, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.1172000840306282, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.1209568977355957, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.11145728826522827, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.10561785101890564, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.09339147061109543, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.05631561577320099, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.05235263332724571, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.07037793099880219, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.06344734877347946, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.05781841278076172, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.05036773905158043, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.046814654022455215, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.03587089106440544, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.03031330183148384, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.027123546227812767, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.026335271075367928, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.017914308235049248, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.014027467928826809, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.01366224605590105, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.012823130004107952, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.011986542493104935, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.009259259328246117, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.009012951515614986, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.007984586991369724, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.005835695657879114, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.1209568977355957, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.1209568977355957, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.10617516189813614, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.09768043458461761, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.09131001681089401, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.0804753303527832, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.04896141216158867, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.044756390154361725, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.06151260808110237, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.05674164742231369, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.05023257061839104, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.04371792823076248, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.040007635951042175, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.031034313142299652, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.027115602046251297, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.02361132577061653, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.022729428485035896, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.01553376019001007, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.012162335216999054, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.011723033152520657, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.011089439503848553, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.010232917964458466, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.008107374422252178, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.007884734310209751, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.006950642913579941, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.004975480958819389, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.10617516189813614, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.10617516189813614, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.26566797494888306, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.24615736305713654, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.23813273012638092, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.21159501373767853, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.12451764196157455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.11798477172851562, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.14731371402740479, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13245552778244019, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.1275089979171753, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.11138850450515747, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.10312386602163315, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07614261656999588, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.06372227519750595, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05976263806223869, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.058879729360342026, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.038245365023612976, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.03016338311135769, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.02980915829539299, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02742054872214794, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.026017028838396072, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.019459469243884087, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.017824355512857437, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.017384512349963188, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010327570140361786, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13245552778244019, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13245552778244019, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.2443016916513443, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.21701647341251373, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.2065630406141281, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.17628240585327148, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11126691102981567, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.10283421725034714, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13192254304885864, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1211007833480835, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.11544570326805115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.09469421207904816, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.08505622297525406, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.06748602539300919, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.05811194330453873, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.053625766187906265, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.0525442473590374, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03388684242963791, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.028096705675125122, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.02770862728357315, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.024810144677758217, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.02330787479877472, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.01812233030796051, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.018194306641817093, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.01666342467069626, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.012584185227751732, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13192254304885864, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13192254304885864, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2314196676015854, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2152264565229416, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.20913292467594147, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.18650367856025696, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.10899730026721954, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10402260720729828, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12433801591396332, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.11488214880228043, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11143898218870163, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.09819906204938889, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.08953677117824554, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06353051960468292, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.055189359933137894, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.052511416375637054, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05187252536416054, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03179936856031418, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.027351589873433113, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.02713475190103054, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.025188900530338287, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.024180084466934204, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.016793644055724144, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01718869060277939, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.015892289578914642, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011815059930086136, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12433801591396332, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12433801591396332, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.28584566712379456, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.26634976267814636, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2593425214290619, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23162946105003357, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13475793600082397, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.12890002131462097, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1530698835849762, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.141117662191391, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13766027987003326, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12147175520658493, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11075762659311295, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07805641740560532, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06756278872489929, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06464354693889618, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06394777446985245, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03899050131440163, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03308379650115967, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03286782279610634, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.030344698578119278, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.029114235192537308, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020242558792233467, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.0198898296803236, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01919686794281006, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012632656842470169, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13766027987003326, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13766027987003326, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.13.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.2997806668281555, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.26608744263648987, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.25369659066200256, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.22346539795398712, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1403946429491043, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12607960402965546, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.1632881462574005, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.15036752820014954, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.14448696374893188, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11739655584096909, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11092512309551239, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08387908339500427, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07251537591218948, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06780048459768295, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06668123602867126, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.042136743664741516, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.035529788583517075, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.03514215350151062, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.030981820076704025, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.03054238110780716, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.02258942648768425, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.022821029648184776, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.021044757217168808, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.015905698761343956, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11739655584096909, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11739655584096909, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.10976319015026093, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.10096851736307144, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.09571803361177444, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.0847306028008461, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.051118142902851105, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.047481559216976166, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.0639839693903923, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.05751894414424896, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.05248960107564926, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.04562915861606598, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.042633891105651855, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.03278721868991852, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.027526531368494034, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.024667471647262573, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.02395542338490486, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.01642768643796444, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.012878731824457645, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.012542292475700378, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.011782974936068058, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.011032615788280964, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.008607480674982071, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.008409490808844566, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.007445999421179295, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.0056330482475459576, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.10976319015026093, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.10976319015026093, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.1063171774148941, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.09750533103942871, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.09153859317302704, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.08058428764343262, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.04883179813623428, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.04496654495596886, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.060915686190128326, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.056140605360269547, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.05026562884449959, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.043623168021440506, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.04004703462123871, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.03068026527762413, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.02673986181616783, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.023531349375844002, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.0226821880787611, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.0154096819460392, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.012035728432238102, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.011634958907961845, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.010962930507957935, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.010156641714274883, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.007921535521745682, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.007695676293224096, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.006846921984106302, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.004701634403318167, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.1063171774148941, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.1063171774148941, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.25720617175102234, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.23710602521896362, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.22871193289756775, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.20266319811344147, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.11981633305549622, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.11298874765634537, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.14372509717941284, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.1288498044013977, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.12295478582382202, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.10712053626775742, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.09886594116687775, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.07394816726446152, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.06194372475147247, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05751222372055054, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.05638275295495987, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.037021178752183914, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.02908986806869507, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.02865566685795784, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02642645314335823, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02496185339987278, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.018878236413002014, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.01745785027742386, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.01667597144842148, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010242690332233906, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.1288498044013977, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.1288498044013977, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.23609237372875214, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.20992812514305115, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.19487179815769196, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.16994987428188324, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.10690616071224213, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.09644539654254913, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.138155996799469, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.1252746284008026, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.11209602653980255, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.09343019127845764, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.08640158176422119, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07079535722732544, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.0603172704577446, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05205904692411423, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.04994923993945122, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.0356760136783123, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.028053831309080124, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.02718830481171608, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.025481760501861572, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.023521654307842255, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.019292492419481277, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.019737716764211655, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.01661316305398941, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.014333884231746197, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.138155996799469, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.138155996799469, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2376617193222046, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2214347869157791, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2153690904378891, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.19240647554397583, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11202608048915863, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10705161839723587, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12771877646446228, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.11790148913860321, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11447851359844208, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10107171535491943, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.0923376977443695, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06528692692518234, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.056676656007766724, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05400683358311653, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05336952582001686, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03269490227103233, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.028143610805273056, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.027921531349420547, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02595025673508644, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.024914758279919624, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017347466200590134, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017703725025057793, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01643766462802887, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.01221905555576086, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12771877646446228, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12771877646446228, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.28320619463920593, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2640817165374756, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.25724437832832336, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23008963465690613, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13342677056789398, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.1277812123298645, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.15185022354125977, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13973122835159302, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1363491415977478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12050474435091019, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11010520160198212, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07750249654054642, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06697382032871246, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06409897655248642, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.0634084939956665, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.038746751844882965, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.032843369990587234, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03262006491422653, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.03016684204339981, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02896573208272457, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.020252862945199013, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01982983574271202, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.019208669662475586, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012710717506706715, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1363491415977478, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1363491415977478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.14.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.3050888180732727, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.27030491828918457, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.25719091296195984, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.22652573883533478, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1430443972349167, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.1281171441078186, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16789235174655914, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.15385843813419342, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.1473444104194641, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11957073956727982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11321450769901276, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08628568798303604, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07442891597747803, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.0693173035979271, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06809955835342407, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.043558306992053986, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03668754920363426, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.036252863705158234, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.03207390010356903, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.031587183475494385, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.023615607991814613, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.02403978444635868, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.021939435973763466, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.01725231483578682, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11957073956727982, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11957073956727982, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.1322113424539566, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.1215248703956604, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.11547262966632843, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.10225245356559753, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.061656784266233444, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.05732342600822449, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.07605426758527756, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.06878410279750824, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.06330893933773041, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.05509486049413681, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.051077231764793396, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.03878229856491089, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.033082664012908936, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.029744258150458336, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.028945695608854294, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.019487254321575165, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.015503223985433578, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.015151762403547764, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.014192725531756878, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.013344667851924896, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.010195373557507992, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.01007765345275402, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.0090103130787611, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.006755854468792677, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.1322113424539566, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.1322113424539566, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.10803651809692383, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.0992317646741867, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.09282849729061127, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.08155256509780884, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.04966351389884949, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.04539462924003601, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.06250502914190292, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.05780978873372078, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.051025547087192535, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.044478561729192734, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.04073212668299675, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.03156021237373352, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.027691900730133057, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.023977186530828476, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.022996025159955025, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.01576688326895237, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.012381167151033878, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.011924803256988525, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.011292608454823494, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.010418007150292397, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.008140885271131992, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.008110199123620987, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.0069396477192640305, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005158070009201765, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.10803651809692383, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.10803651809692383, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.26629146933555603, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.24505704641342163, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.23480236530303955, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.20811037719249725, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.12407414615154266, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.1162298247218132, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.15076947212219238, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13543173670768738, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.12758341431617737, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.11072559654712677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.10322283208370209, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.0782586932182312, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.06520869582891464, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.05957851931452751, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.05825043097138405, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.039299171417951584, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.030273281037807465, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.029652860015630722, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.027383912354707718, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02578471042215824, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.020030999556183815, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.018319839611649513, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.017416296526789665, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.010742942802608013, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13543173670768738, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.13543173670768738, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.2434520423412323, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.2210671603679657, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.21289175748825073, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.18601596355438232, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11139054596424103, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.10475129634141922, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13003306090831757, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.11904911696910858, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.11450515687465668, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.09796269983053207, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.08848890662193298, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.06655018031597137, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.05714983493089676, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05362113565206528, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.05279030278325081, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03332718834280968, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.02793322689831257, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.0276167094707489, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.025245612487196922, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.023996757343411446, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.017660517245531082, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.017680158838629723, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.016489410772919655, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.012050674296915531, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13003306090831757, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13003306090831757, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.24311180412769318, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.22627034783363342, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.21997973322868347, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.19649572670459747, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11441885679960251, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10927741974592209, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.130562886595726, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12052777409553528, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11699498444795609, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.1032312735915184, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09415114670991898, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06654047220945358, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05780494585633278, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05502171814441681, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05436267331242561, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03326619043946266, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.028428999707102776, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.028204403817653656, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.026146570220589638, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.025086674839258194, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017423629760742188, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01756284199655056, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.016473820433020592, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011704504489898682, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.130562886595726, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.130562886595726, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2841663658618927, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.26471689343452454, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2577945590019226, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.23043610155582428, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13379548490047455, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.12801578640937805, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.15207739174365997, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.14021195471286774, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1367453932762146, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.12071074545383453, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.11009299755096436, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07740712910890579, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06707245111465454, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06413166970014572, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06344009190797806, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03866391256451607, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.032726921141147614, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03250565752387047, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.030014969408512115, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02878243662416935, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.019968358799815178, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01954662799835205, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.018901020288467407, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012219613417983055, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1367453932762146, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1367453932762146, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.15.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.3003634214401245, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.2656846344470978, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2509625554084778, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.22103706002235413, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.14072257280349731, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.1250067949295044, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.1685316413640976, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.15415701270103455, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.14516189694404602, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11768212169408798, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11160966008901596, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08653473109006882, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.0743900015950203, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06836897134780884, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06688442081212997, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.043563034385442734, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03641815483570099, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.035854920744895935, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.03195822611451149, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.03131118416786194, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.02365564927458763, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.02438088320195675, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.021685196086764336, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.017727378755807877, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11768212169408798, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11768212169408798, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.12936173379421234, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.11875467747449875, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.11251920461654663, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.09967505931854248, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.060089271515607834, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.05570314824581146, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.0758737176656723, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.06779937446117401, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.06197025626897812, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.05386540666222572, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.050172626972198486, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.038799673318862915, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.03257855027914047, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.029018498957157135, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.02812804840505123, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.01936308853328228, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.015108983963727951, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.014714418910443783, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.013849431648850441, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.012952856719493866, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.010145578533411026, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.009870076552033424, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.00874456949532032, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.006576229818165302, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.12936173379421234, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.12936173379421234, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.11156586557626724, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.1027488261461258, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.09563282132148743, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.08479925245046616, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.05160364881157875, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.047050297260284424, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.0650654062628746, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.060897063463926315, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.05306678265333176, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.046279728412628174, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.04255671799182892, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.0330672524869442, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.029109304770827293, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.024881433695554733, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.023784514516592026, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.016525298357009888, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.012822183780372143, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.01225526537746191, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.011737644672393799, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.010721336118876934, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.008653209544718266, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.00841980054974556, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.007313479203730822, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005177539773285389, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11156586557626724, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11156586557626724, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.288030207157135, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.266899436712265, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.2577062249183655, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.22892190515995026, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.13511967658996582, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.12790806591510773, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.15908384323120117, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.14453451335430145, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.13836142420768738, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.12122383713722229, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.1111830621957779, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.0820963904261589, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.06939784437417984, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.06466007232666016, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.0634879618883133, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.04141673818230629, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.032692600041627884, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.03216226398944855, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.02973233163356781, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.028126206248998642, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.021008554846048355, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.01940818503499031, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.018757347017526627, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.011148277670145035, "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": 3.193359375, "total_bits": 1674240.0, "err": 0.13836142420768738, "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": 3.193359375, "total_bits": 1674240.0, "err": 0.13836142420768738, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.16.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.25058552622795105, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.2276799976825714, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.21922831237316132, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.18812590837478638, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11554113030433655, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.1087188869714737, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13478316366672516, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12285421788692474, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.11884955316781998, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10048232972621918, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.08916870504617691, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.0690685287117958, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.059034544974565506, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.055635832250118256, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.054811280220746994, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03453202173113823, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.02894420176744461, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.028671978041529655, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.02591877430677414, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.024557866156101227, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.018303008750081062, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.01828470453619957, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.017182687297463417, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.012524357996881008, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13478316366672516, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.13478316366672516, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.23336157202720642, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2173355221748352, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.21135082840919495, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.18880252540111542, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.10980295389890671, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.10492721945047379, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12488128989934921, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.11557453870773315, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11227157711982727, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.09909247606992722, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09037782996892929, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06368638575077057, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05540978163480759, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05279364064335823, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05216745659708977, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03181992098689079, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.027249831706285477, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.027047643437981606, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.025071680545806885, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02405615523457527, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01663215085864067, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.016780825331807137, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.015755098313093185, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011133858934044838, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12488128989934921, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.12488128989934921, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.27683913707733154, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2581132650375366, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2515144348144531, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.22488872706890106, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13030622899532318, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.12473408132791519, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.14769816398620605, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13635121285915375, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1331409066915512, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.11767245084047318, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.10730992257595062, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07520899921655655, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06524878740310669, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06248106807470322, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.061809245496988297, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.037533070892095566, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03184723109006882, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.03164016455411911, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.029236627742648125, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02805090881884098, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.019353758543729782, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.018963199108839035, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01835975982248783, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011809278279542923, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13635121285915375, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13635121285915375, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.16.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.2946954369544983, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.2598125636577606, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.24594813585281372, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.21645577251911163, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13789136707782745, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12252312153577805, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16392768919467926, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.14983251690864563, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.142191082239151, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11481229215860367, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.10888875275850296, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08411888033151627, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07236914336681366, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.0669780969619751, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.0656510442495346, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.04231967031955719, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.035608578473329544, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.035135481506586075, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.031133439391851425, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.030576221644878387, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.02289208583533764, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.023685913532972336, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.02112129144370556, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.01722416654229164, "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": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13789136707782745, "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": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13789136707782745, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.12438561767339706, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.11471010744571686, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.10869155079126358, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.09605380892753601, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.057926371693611145, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.05382025986909866, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.07202505320310593, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.06555783748626709, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.05947563424706459, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.051802534610033035, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.047882404178380966, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.0364999994635582, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.03131493553519249, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.027977032586932182, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.027122993022203445, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.018269073218107224, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.014540263451635838, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.01413393672555685, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.013303836807608604, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.012430492788553238, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.009608189575374126, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.009472332894802094, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.008469375781714916, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.006232416722923517, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.12438561767339706, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.12438561767339706, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.11310158669948578, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.10432827472686768, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.0967220664024353, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.08569064736366272, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.05211358144879341, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.04743613302707672, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.0667169913649559, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.061980340629816055, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.05361279845237732, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.0469231940805912, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.04336311295628548, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.03334968164563179, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.02943740040063858, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.02512037754058838, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.02401585318148136, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.016624784097075462, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.01290508359670639, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.012329580262303352, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.011814968660473824, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.01079021580517292, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.008649159222841263, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.00840933620929718, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.007339468225836754, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005094033665955067, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11310158669948578, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11310158669948578, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.3086055517196655, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.2839401960372925, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.27345582842826843, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.2412118762731552, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.14391562342643738, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.13535583019256592, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.17019888758659363, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.15465803444385529, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.1481296420097351, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.12824851274490356, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.11787360906600952, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.08870230615139008, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.07469979673624039, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.06919237971305847, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.06777060776948929, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.044939760118722916, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.03517845273017883, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.03461536392569542, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.03185095638036728, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.030124835669994354, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.023088296875357628, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.02137714996933937, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.02020275965332985, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.01287832111120224, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12824851274490356, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12824851274490356, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.17.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.28642839193344116, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.26448073983192444, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.25691258907318115, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.22520843148231506, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.13296383619308472, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.1266043484210968, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.15295933187007904, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.13970257341861725, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.13622388243675232, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.11843010038137436, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.10688669979572296, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07837105542421341, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.06709831953048706, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.06396802514791489, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.06321391463279724, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03922285512089729, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.03318006917834282, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.03296355530619621, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.030245110392570496, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.028965309262275696, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.02072795480489731, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.02068944089114666, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.019709881395101547, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.014102943241596222, "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": 3.188232421875, "total_bits": 13372416.0, "err": 0.13622388243675232, "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": 3.188232421875, "total_bits": 13372416.0, "err": 0.13622388243675232, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.24711595475673676, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.230171337723732, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.2239995300769806, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.2001490592956543, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11637910455465317, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11124415695667267, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13199065625667572, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12214925140142441, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11892422288656235, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10501386225223541, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.0957367792725563, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06732864677906036, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05858195945620537, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.0559220127761364, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.055290065705776215, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.033649176359176636, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.02883191406726837, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.0286247618496418, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02650654874742031, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.025450920686125755, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017593415454030037, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017674364149570465, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01670716516673565, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011689247563481331, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13199065625667572, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13199065625667572, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.2821371555328369, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2628408372402191, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.25608140230178833, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.22884772717952728, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.13278642296791077, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.1271417737007141, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.15053269267082214, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1389905959367752, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13568910956382751, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.1198652908205986, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.10931921005249023, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07667263597249985, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.0664757713675499, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06368046998977661, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06301075220108032, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03827629238367081, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.0324433408677578, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.032232630997896194, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.029739152640104294, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02854294143617153, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.019767774268984795, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.019290385767817497, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.01876971125602722, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011977353133261204, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1389905959367752, "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": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13568910956382751, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.17.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.3066721260547638, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.2716580927371979, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2583580017089844, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.22770744562149048, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.1435108482837677, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12844234704971313, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16861820220947266, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.15444473922252655, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.1477939635515213, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.12000870704650879, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.11351852864027023, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08636119216680527, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.07438261806964874, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06936737895011902, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.06815406680107117, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.043343741446733475, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.036351919174194336, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.035916924476623535, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.03169496729969978, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.031202200800180435, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.023264339193701744, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.023385092616081238, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.02159780263900757, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.016276201233267784, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.12000870704650879, "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": 3.534823330965909, "total_bits": 40771840.0, "err": 0.12000870704650879, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.13945066928863525, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.12806612253189087, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.12133288383483887, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.1077260747551918, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.06516262888908386, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.06022144854068756, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.08038849383592606, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.07328566908836365, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.06691423058509827, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.05818864330649376, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.053710002452135086, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.04095032811164856, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.03519030287861824, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.031469594687223434, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.030525660142302513, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.02053125761449337, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.01635998860001564, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.015939021483063698, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.014954722486436367, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.013992922380566597, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.010797630995512009, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.010639671236276627, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.009554796852171421, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.007038661744445562, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13945066928863525, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13945066928863525, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.18.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.1166725680232048, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.10757507383823395, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.09945037961006165, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.08825382590293884, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.05405152589082718, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.04880940914154053, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.06955135613679886, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.06455092877149582, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.05547745153307915, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.048452552407979965, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.04466133564710617, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.03484143689274788, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.03075406327843666, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.02604215033352375, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.024753334000706673, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.017463738098740578, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.013452084735035896, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.01284036599099636, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.012330499477684498, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.011252952739596367, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.009107435122132301, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.008943893015384674, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.007743642199784517, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005611647851765156, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.1166725680232048, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.1166725680232048, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.30280447006225586, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.275913268327713, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.26354002952575684, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.23158955574035645, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.14082694053649902, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.13080573081970215, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.17098261415958405, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.1543741375207901, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.14539816975593567, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.12454763054847717, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.1154462993144989, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.08889564871788025, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.07442222535610199, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.06760954856872559, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.06583910435438156, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.04510612040758133, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.03436935693025589, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.03362775221467018, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.030979687348008156, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02887456677854061, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.022997234016656876, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.021106675267219543, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.0197150856256485, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.012452286668121815, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12454763054847717, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12454763054847717, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.25407111644744873, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.22942984104156494, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.22065642476081848, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.19104541838169098, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.11620937287807465, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.10901840776205063, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.1367785632610321, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12484076619148254, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12031946331262589, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10153304785490036, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.09250425547361374, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.0698738768696785, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.059795524924993515, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.0558861680328846, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.05492405965924263, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03493808954954147, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.028985872864723206, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.028685055673122406, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.02601456642150879, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.02457835152745247, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.01839642971754074, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.018255136907100677, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.017058491706848145, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.012245628982782364, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.1367785632610321, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.1367785632610321, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.24542953073978424, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.228555366396904, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.22243623435497284, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.19865453243255615, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11558373272418976, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11056587100028992, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13105596601963043, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1212441697716713, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.11820383369922638, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.1041952446103096, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09501408040523529, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06687530875205994, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05814484879374504, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.055585309863090515, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.054952483624219894, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.033460769802331924, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.028692202642560005, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.028505083173513412, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.026370495557785034, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02532276324927807, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017565341666340828, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01763479970395565, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.016714276745915413, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011717824265360832, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13105596601963043, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13105596601963043, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.27754148840904236, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.25865447521209717, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.25208672881126404, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.22517333924770355, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1306910663843155, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.12514130771160126, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1476028859615326, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13655352592468262, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.1335107535123825, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.11781667172908783, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.10737915337085724, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0751633271574974, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06530750542879105, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.0626128688454628, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.06197991967201233, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.037495456635951996, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.031888045370578766, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.031689830124378204, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02921561524271965, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.0280429869890213, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.019320880994200706, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.018914517015218735, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.018388211727142334, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011702342890202999, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13655352592468262, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.13655352592468262, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.18.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.2948685586452484, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.26039785146713257, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2474498748779297, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.21789319813251495, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13774579763412476, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.12292458862066269, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.16170695424079895, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.1486421823501587, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.14210890233516693, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11505014449357986, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.1089540645480156, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08300773799419403, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.0716894119977951, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06659854203462601, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.0653490275144577, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.041766099631786346, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03493638336658478, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.034493338316679, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.030446147546172142, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.029931388795375824, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.022345442324876785, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.022571967914700508, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.020643556490540504, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.01574569195508957, "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": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13774579763412476, "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": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13774579763412476, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.13579072058200836, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.12487665563821793, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.11807550489902496, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.10487537086009979, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.06323783099651337, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.058527909219264984, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.07892928272485733, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.0717652291059494, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.06518205255270004, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.056731853634119034, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.05246372148394585, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.040044814348220825, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.03439580649137497, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.030536644160747528, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.02957073412835598, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.02007957547903061, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.015930689871311188, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.015492587350308895, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.014625674113631248, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.0136564327403903, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.010625234805047512, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.010437841527163982, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.009317375719547272, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.00692770117893815, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13579072058200836, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13579072058200836, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.19.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.11752711236476898, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.10794715583324432, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.100297711789608, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.08895044028759003, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.05437323823571205, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.04936781898140907, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.0690249428153038, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.06423762440681458, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.055987145751714706, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.048634689301252365, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.044818051159381866, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.03473616763949394, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.030686473473906517, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.02611115388572216, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.02495993673801422, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.01737169548869133, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.013449297286570072, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.012887497432529926, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.012294843792915344, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.011233842931687832, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.009001035243272781, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.008809641934931278, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.007634900510311127, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005373872350901365, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11752711236476898, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11752711236476898, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.3073562681674957, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.2823239266872406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.27144917845726013, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.24012896418571472, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.14396882057189941, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.13520890474319458, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.17167803645133972, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.15492555499076843, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.14805783331394196, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.12779906392097473, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.11693111807107925, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.08846087753772736, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.0743599534034729, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.06902731209993362, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.06763733178377151, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.04439454525709152, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.03505155071616173, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.03458160161972046, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.03171931579709053, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.02988852746784687, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.02275693789124489, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.02115762047469616, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.02019049972295761, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.012597125954926014, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12779906392097473, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12779906392097473, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.2631568908691406, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.24202755093574524, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.23449814319610596, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.20588399469852448, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.12207215279340744, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.11565986275672913, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.1414100080728531, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12923631072044373, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.1253722459077835, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10838525742292404, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.09759373962879181, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.0724763348698616, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.0620468407869339, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.05877603590488434, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.05793999508023262, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03621993958950043, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.030389264225959778, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.030131658539175987, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.027634503319859505, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.0263358186930418, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.01898801513016224, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.018928909674286842, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.01788286492228508, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.01271804515272379, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12923631072044373, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12923631072044373, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.25173524022102356, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.23431947827339172, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.22811147570610046, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.203862726688385, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11861121654510498, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11341153830289841, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13425710797309875, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12438131123781204, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12128198891878128, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10688026994466782, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09746262431144714, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.0684562623500824, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.059571631252765656, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05696435645222664, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.056329965591430664, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03421109914779663, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.029254401102662086, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.029064010828733444, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.026850871741771698, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02578319050371647, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.017807120457291603, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017778612673282623, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.016925977542996407, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011535214260220528, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13425710797309875, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.13425710797309875, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.27422916889190674, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.2553103268146515, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.24878422915935516, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.22228007018566132, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.1290888637304306, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.12363633513450623, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.14570419490337372, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.134924054145813, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.13190694153308868, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.11639612168073654, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.10600024461746216, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07414176315069199, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.0645073801279068, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.06184421479701996, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.061221472918987274, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03700900450348854, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.03149024769663811, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.0312979593873024, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02885233797132969, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.027688317000865936, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.019048992544412613, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.018676260486245155, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.018137091770768166, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011518817394971848, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.134924054145813, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.134924054145813, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.19.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.28587767481803894, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.2535551190376282, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.2411058247089386, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.21194249391555786, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.13366290926933289, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.11958988010883331, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.156902015209198, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.14423802495002747, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13768628239631653, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.11183156073093414, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.1056494265794754, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.08052986860275269, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.06944064795970917, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.06453099846839905, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.0633450299501419, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.04051605984568596, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.03364791348576546, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.033204153180122375, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.029275933280587196, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.028772709891200066, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.021765854209661484, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.0214788019657135, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.02014760486781597, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.014635429717600346, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13768628239631653, "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": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.13768628239631653, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.13021884858608246, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.12081190943717957, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.11502432823181152, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.10249633342027664, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.061225876212120056, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.05713486298918724, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.07561570405960083, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.06821097433567047, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.06271632015705109, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.055132921785116196, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.051391199231147766, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.03871519863605499, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.0328032486140728, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.029535694047808647, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.028702901676297188, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.019469816237688065, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.015287919901311398, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.014919258654117584, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.014045270159840584, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.013202094472944736, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.010248392820358276, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.009796541184186935, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.00904197245836258, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.006396280601620674, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13021884858608246, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13021884858608246, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.11870184540748596, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.10936249792575836, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.10236028581857681, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.09085141867399216, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.055182263255119324, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.050533682107925415, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.06861954927444458, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.06391999125480652, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.0566057525575161, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.049480460584163666, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.04542040452361107, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.03455324470996857, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.03054259903728962, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.026521630585193634, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.02543761394917965, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.017317673191428185, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.013527260161936283, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.01301498617976904, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.012349541299045086, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.011359610594809055, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.008943408727645874, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.008646231144666672, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.0077091967687010765, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.005159509833902121, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11870184540748596, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.11870184540748596, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.30099329352378845, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.27814981341362, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.2701805830001831, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.23917324841022491, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.14061537384986877, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.13385887444019318, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.16279703378677368, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.14828361570835114, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.1442279815673828, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.12564332783222198, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.11469589173793793, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.0836620181798935, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.07099901139736176, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.06736142933368683, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.06648855656385422, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.041785381734371185, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.034004759043455124, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.03373488783836365, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.030787035822868347, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.029382461681962013, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.021304944530129433, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.019912494346499443, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.01958412304520607, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.0115520553663373, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12564332783222198, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12564332783222198, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } } }, { "key": "model.layers.20.self_attn.o_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.25796064734458923, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.23364956676959991, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.2254294902086258, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.19646844267845154, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.1193234771490097, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.11239548772573471, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.138187438249588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.12670621275901794, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.12316573411226273, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.10445733368396759, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.09387434273958206, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.07099274545907974, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.06098322570323944, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.0576135590672493, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.05679957568645477, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.03554798290133476, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.030040763318538666, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.02980107255280018, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.027035264298319817, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.025709113106131554, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.018836669623851776, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.019010106101632118, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.017772652208805084, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.013087422586977482, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.138187438249588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.138187438249588, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.25333037972450256, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.23592382669448853, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.22963520884513855, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.20521633327007294, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.11931890994310379, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11408080160617828, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1350383311510086, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12511274218559265, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12202627211809158, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.10756612569093704, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.09804823994636536, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.068873830139637, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05992887169122696, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.057280298322439194, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.05665324255824089, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.03442038223147392, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.029415205121040344, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.02922207862138748, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02699984423816204, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.025924142450094223, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01793091371655464, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.017847532406449318, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.017039252445101738, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.011576175689697266, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1350383311510086, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.1350383311510086, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.26372334361076355, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.24562430381774902, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.23949271440505981, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.21396905183792114, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.12412454932928085, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.11883485317230225, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.14022177457809448, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12972527742385864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.12687472999095917, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.11197972297668457, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.10196942836046219, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.07134153693914413, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.06206944212317467, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.05951634421944618, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.0589054599404335, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.035628560930490494, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.030294299125671387, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.030113155022263527, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02775699272751808, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.026640448719263077, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01833452098071575, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.01796703226864338, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.017450960353016853, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.01110281702131033, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12972527742385864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.12972527742385864, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.20.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.2604832649230957, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.23102693259716034, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.21955078840255737, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.19349734485149384, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.12165165692567825, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.10878489911556244, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.1440584361553192, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.13172955811023712, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.12542450428009033, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.10214364528656006, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.09675472974777222, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.073858343064785, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.06362639367580414, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.05888785049319267, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.05775361508131027, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.037260595709085464, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.031029891222715378, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.030611464753746986, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.027184724807739258, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.02670588716864586, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.020183466374874115, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.020250195637345314, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.018616467714309692, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.01431155763566494, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.13172955811023712, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.13172955811023712, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.q_proj", "numel": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.13967324793338776, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.12913604080677032, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.12259828299283981, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.10884230583906174, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.06539983302354813, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.06083337590098381, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.08064466714859009, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.07354652881622314, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.06707405298948288, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.058589398860931396, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.054074592888355255, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.04102657735347748, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.03530241176486015, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.03160060942173004, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.030693404376506805, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.020536154508590698, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.01646294817328453, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.016037749126553535, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.015094947069883347, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.014134719967842102, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.010840244591236115, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.010725517757236958, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.009585035964846611, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.007150785066187382, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13967324793338776, "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": 2.188232421875, "total_bits": 9178112.0, "err": 0.13967324793338776, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } } }, { "key": "model.layers.21.self_attn.k_proj", "numel": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.13219793140888214, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.12177994102239609, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.1149025484919548, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.10222023725509644, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.06143159419298172, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.05676145479083061, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.07570400089025497, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.06990054994821548, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.06300872564315796, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.054884374141693115, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.050652798265218735, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.03807581588625908, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.0334068201482296, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.029555123299360275, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.028572773560881615, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.019108640030026436, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.015210296958684921, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.01475929468870163, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.013883496634662151, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.012940851040184498, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.010022932663559914, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.009737513028085232, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.008825426921248436, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.0061216093599796295, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.13219793140888214, "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": 2.193359375, "total_bits": 1149952.0, "err": 0.13219793140888214, "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": 524288, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.193359375, "total_bits": 1149952.0, "err": 0.30503833293914795, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.380859375, "total_bits": 1248256.0, "err": 0.281886488199234, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.630859375, "total_bits": 1379328.0, "err": 0.27357131242752075, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.755859375, "total_bits": 1444864.0, "err": 0.2421448975801468, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.255859375, "total_bits": 1707008.0, "err": 0.1428472101688385, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.693359375, "total_bits": 1936384.0, "err": 0.13609981536865234, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03271484375, "total_bits": 1590016.0, "err": 0.16462627053260803, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.130859375, "total_bits": 1641472.0, "err": 0.15028776228427887, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.193359375, "total_bits": 1674240.0, "err": 0.14637775719165802, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.505859375, "total_bits": 1838080.0, "err": 0.12757514417171478, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6904296875, "total_bits": 1934848.0, "err": 0.11621750891208649, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03271484375, "total_bits": 2114304.0, "err": 0.08430919051170349, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.130859375, "total_bits": 2165760.0, "err": 0.0719262883067131, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.255859375, "total_bits": 2231296.0, "err": 0.0685303583741188, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.380859375, "total_bits": 2296832.0, "err": 0.06771191209554672, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03271484375, "total_bits": 2638592.0, "err": 0.04214738309383392, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.255859375, "total_bits": 2755584.0, "err": 0.03462786599993706, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.380859375, "total_bits": 2821120.0, "err": 0.03437887877225876, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.505859375, "total_bits": 2886656.0, "err": 0.03140442818403244, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.818359375, "total_bits": 3050496.0, "err": 0.029955368489027023, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03271484375, "total_bits": 3162880.0, "err": 0.021500878036022186, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.130859375, "total_bits": 3214336.0, "err": 0.020180484279990196, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28271484375, "total_bits": 3293952.0, "err": 0.019961731508374214, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.130859375, "total_bits": 4262912.0, "err": 0.011713100597262383, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12757514417171478, "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": 3.505859375, "total_bits": 1838080.0, "err": 0.12757514417171478, "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": 4194304, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.188232421875, "total_bits": 9178112.0, "err": 0.16482388973236084, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 9964544.0, "err": 0.1461755484342575, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 11013120.0, "err": 0.14037515223026276, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750732421875, "total_bits": 11537408.0, "err": 0.12079520523548126, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250732421875, "total_bits": 13634560.0, "err": 0.0755574107170105, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.688232421875, "total_bits": 15469568.0, "err": 0.07061133533716202, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 12714752.0, "err": 0.08781952410936356, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 13110272.0, "err": 0.08024463802576065, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.188232421875, "total_bits": 13372416.0, "err": 0.07818800956010818, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500732421875, "total_bits": 14683136.0, "err": 0.06419473886489868, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6878662109375, "total_bits": 15468032.0, "err": 0.05788510665297508, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 16909056.0, "err": 0.045086972415447235, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 17304576.0, "err": 0.03866887465119362, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.250732421875, "total_bits": 17828864.0, "err": 0.03657394275069237, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.375732421875, "total_bits": 18353152.0, "err": 0.03607102483510971, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 21103360.0, "err": 0.022631559520959854, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.250732421875, "total_bits": 22023168.0, "err": 0.019365506246685982, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.375732421875, "total_bits": 22547456.0, "err": 0.019227037206292152, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.500732421875, "total_bits": 23071744.0, "err": 0.017157698050141335, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.813232421875, "total_bits": 24382464.0, "err": 0.016382234171032906, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 25297664.0, "err": 0.012259082868695259, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 25693184.0, "err": 0.01267695240676403, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.28143310546875, "total_bits": 26346240.0, "err": 0.011581790633499622, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 34081792.0, "err": 0.009230035357177258, "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": 2.750732421875, "total_bits": 11537408.0, "err": 0.12079520523548126, "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": 2.750732421875, "total_bits": 11537408.0, "err": 0.12079520523548126, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.gate_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.21973641216754913, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.20423440635204315, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.19875304400920868, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.17771375179290771, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.10393671691417694, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.0993318036198616, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.11825832724571228, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.1091461256146431, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.10627942532300949, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.09375345706939697, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.0857226625084877, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.06104379892349243, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.05299566313624382, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.050609320402145386, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.050028569996356964, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.030798794701695442, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.027261927723884583, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.02709215320646763, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.02532282844185829, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.024433474987745285, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.01715618185698986, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.0182871762663126, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.016379550099372864, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.013956304639577866, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.11825832724571228, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.11825832724571228, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.up_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.187766335227273, "total_bits": 25234432.000000004, "err": 0.1847129464149475, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375266335227273, "total_bits": 27397120.000000004, "err": 0.17168742418289185, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625266335227273, "total_bits": 30280704.000000004, "err": 0.16721944510936737, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.750266335227273, "total_bits": 31722496.000000004, "err": 0.1494114100933075, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.250266335227273, "total_bits": 37489664.0, "err": 0.08748218417167664, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.687766335227273, "total_bits": 42535936.0, "err": 0.08368869125843048, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.09936706721782684, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125266335227273, "total_bits": 36047872.0, "err": 0.09169090539216995, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.187766335227273, "total_bits": 36768768.0, "err": 0.08957581967115402, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.500266335227273, "total_bits": 40373248.0, "err": 0.0789775550365448, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.6876331676136362, "total_bits": 42534400.0, "err": 0.07227756828069687, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.031316583806818, "total_bits": 46498560.0, "err": 0.051470253616571426, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.1252663352272725, "total_bits": 47582208.0, "err": 0.04476142302155495, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.2502663352272725, "total_bits": 49024000.0, "err": 0.04281187430024147, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.3752663352272725, "total_bits": 50465792.0, "err": 0.04235762730240822, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.031316583806818, "total_bits": 58032896.0, "err": 0.026025842875242233, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.2502663352272725, "total_bits": 60558336.0, "err": 0.023363646119832993, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.3752663352272725, "total_bits": 62000128.0, "err": 0.023239534348249435, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.5002663352272725, "total_bits": 63441920.0, "err": 0.021750561892986298, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.8127663352272725, "total_bits": 67046400.0, "err": 0.02103707566857338, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.031316583806818, "total_bits": 69567232.0, "err": 0.014639683067798615, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.1252663352272725, "total_bits": 70650880.0, "err": 0.016020184382796288, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.281316583806818, "total_bits": 72450816.0, "err": 0.014003226533532143, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125266335227273, "total_bits": 93719552.0, "err": 0.012571237981319427, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.09936706721782684, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.0313165838068183, "total_bits": 34964224.0, "err": 0.09936706721782684, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } }, { "key": "model.layers.21.mlp.down_proj", "numel": 11534336, "options": [ { "desc": "0.05:3b/0.95:2b 32g s4", "bpw": 2.1711869673295454, "total_bits": 25043200.0, "err": 0.20318390429019928, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.25:3b/0.75:2b 32g s4", "bpw": 2.375732421875, "total_bits": 27402496.0, "err": 0.1827220469713211, "qparams": { "group_size": 32, "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.25:4b/0.75:2b 32g s4", "bpw": 2.625732421875, "total_bits": 30286080.0, "err": 0.17277564108371735, "qparams": { "group_size": 32, "bits": [ 4, 2 ], "bits_prop": [ 0.25, 0.75 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", "bpw": 2.716641512784091, "total_bits": 31334656.0, "err": 0.15230345726013184, "qparams": { "group_size": 32, "bits": [ 4, 3, 2 ], "bits_prop": [ 0.1, 0.4, 0.5 ], "scale_bits": 4 } }, { "desc": "0.1:4b/0.9:3b 32g s4", "bpw": 3.216641512784091, "total_bits": 37101824.0, "err": 0.095155730843544, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.2:6b/0.8:3b 32g s4", "bpw": 3.7393687855113638, "total_bits": 43131136.0, "err": 0.08577878028154373, "qparams": { "group_size": 32, "bits": [ 6, 3 ], "bits_prop": [ 0.2, 0.8 ], "scale_bits": 4 } }, { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.11890333145856857, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:3b 32g s4", "bpw": 3.125732421875, "total_bits": 36053248.0, "err": 0.10546057671308517, "qparams": { "group_size": 32, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.05:4b/0.95:3b 32g s4", "bpw": 3.1711869673295454, "total_bits": 36577536.0, "err": 0.09799709171056747, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.05, 0.95 ], "scale_bits": 4 } }, { "desc": "0.4:4b/0.6:3b 32g s4", "bpw": 3.534823330965909, "total_bits": 40771840.0, "err": 0.08155401051044464, "qparams": { "group_size": 32, "bits": [ 4, 3 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.6:4b/0.4:3b 64g s4", "bpw": 3.653775301846591, "total_bits": 42143872.0, "err": 0.07780656963586807, "qparams": { "group_size": 64, "bits": [ 4, 3 ], "bits_prop": [ 0.6, 0.4 ], "scale_bits": 4 } }, { "desc": "1.0:4b 128g s4", "bpw": 4.03143310546875, "total_bits": 46499904.0, "err": 0.06057605892419815, "qparams": { "group_size": 128, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:4b 32g s4", "bpw": 4.125732421875, "total_bits": 47587584.0, "err": 0.0515625923871994, "qparams": { "group_size": 32, "bits": [ 4 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:5b/0.9:4b 32g s4", "bpw": 4.216641512784091, "total_bits": 48636160.0, "err": 0.04682699963450432, "qparams": { "group_size": 32, "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:4b 32g s4", "bpw": 4.307550603693182, "total_bits": 49684736.0, "err": 0.04557603597640991, "qparams": { "group_size": 32, "bits": [ 6, 4 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:5b 128g s4", "bpw": 5.03143310546875, "total_bits": 58034240.0, "err": 0.031006991863250732, "qparams": { "group_size": 128, "bits": [ 5 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:6b/0.9:5b 32g s4", "bpw": 5.216641512784091, "total_bits": 60170496.0, "err": 0.025756238028407097, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", "bpw": 5.307550603693182, "total_bits": 61219072.0, "err": 0.025197986513376236, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.05, 0.05, 0.9 ], "scale_bits": 4 } }, { "desc": "0.4:6b/0.6:5b 32g s4", "bpw": 5.534823330965909, "total_bits": 63840512.0, "err": 0.0232382919639349, "qparams": { "group_size": 32, "bits": [ 6, 5 ], "bits_prop": [ 0.4, 0.6 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", "bpw": 5.693914240056818, "total_bits": 65675520.0, "err": 0.022633859887719154, "qparams": { "group_size": 32, "bits": [ 8, 6, 5 ], "bits_prop": [ 0.1, 0.3, 0.6 ], "scale_bits": 4 } }, { "desc": "1.0:6b 128g s4", "bpw": 6.03143310546875, "total_bits": 69568576.0, "err": 0.01754428818821907, "qparams": { "group_size": 128, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "1.0:6b 32g s4", "bpw": 6.125732421875, "total_bits": 70656256.0, "err": 0.018296143040060997, "qparams": { "group_size": 32, "bits": [ 6 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, { "desc": "0.1:8b/0.9:6b 128g s4", "bpw": 6.213251287286932, "total_bits": 71665728.0, "err": 0.015791572630405426, "qparams": { "group_size": 128, "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "scale_bits": 4 } }, { "desc": "1.0:8b 32g s4", "bpw": 8.125732421875, "total_bits": 93724928.0, "err": 0.014112234115600586, "qparams": { "group_size": 32, "bits": [ 8 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } ], "best_option_max": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.11890333145856857, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } }, "best_option": { "desc": "1.0:3b 128g s4", "bpw": 3.03143310546875, "total_bits": 34965568.0, "err": 0.11890333145856857, "qparams": { "group_size": 128, "bits": [ 3 ], "bits_prop": [ 1.0 ], "scale_bits": 4 } } } ], "last_module_idx": 46, "base_perplexity": 12.712201576267528 }