diff --git "a/job.json" "b/job.json" new file mode 100644--- /dev/null +++ "b/job.json" @@ -0,0 +1,149369 @@ +{ + "in_dir": "/mnt/models/jondurbin_airoboros-c34b-2.2.1", + "out_dir": "/mnt/models/jondurbin_airoboros-c34b-2.2.1-4.65bpw", + "cal_dataset": "/opt/wikitext.parquet", + "dataset_rows": 100, + "measurement_rows": 16, + "gpu_rows": 0, + "length": 2048, + "measurement_length": 2048, + "bits": 4.65, + "head_bits": 6, + "progress": "finished", + "shard_size": 8192, + "cal_filename": "/mnt/models/jondurbin_airoboros-c34b-2.2.1-4.65bpw/cal_data.safetensors", + "last_module_idx": 98, + "measurement": [ + { + "key": "model.layers.0.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.009754983708262444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.009649151936173439, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.004874398931860924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0048429472371935844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.004819005262106657, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.002503378316760063, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.015137135982513428, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.009563332423567772, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.004827974364161491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.004800860304385424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.005100698210299015, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0056840539909899235, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0047936756163835526, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.00305476738139987, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.002485387260094285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0031636529602110386, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.002478445880115032, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0022838618606328964, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0024773553013801575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.002282205270603299, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0024034602101892233, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.002476863795891404, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0021275749895721674, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.002281338442116976, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.009754983708262444, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.012308869510889053, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0121744554489851, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.005809352267533541, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.005770930089056492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.005744478665292263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0019422179320827127, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.013868228532373905, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.012074759230017662, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.005758374929428101, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.005721075925976038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.006120192818343639, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.00633599329739809, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.005713554564863443, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0030332552269101143, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0019125466933473945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0032317056320607662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0018991060787811875, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0014158766716718674, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0018967142095789313, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0014117771061137319, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0018490661168470979, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0018961172318086028, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0011761040659621358, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.001409206772223115, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.012308869510889053, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06038277596235275, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04954804852604866, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03892793506383896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03230416029691696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02663472294807434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.018685800954699516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04966844245791435, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03815614432096481, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02822214365005493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02119661681354046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02155807428061962, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.025685228407382965, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01834719069302082, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013096318580210209, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011532355099916458, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.012800007127225399, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0071880710311234, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006444950588047504, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006205669138580561, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005244108848273754, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006524601485580206, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005731028039008379, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0037105954252183437, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003967223223298788, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.03230416029691696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.010789825581014156, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.007634199224412441, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.006059742532670498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.005612291861325502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.004337919410318136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0030563168693333864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.006624551955610514, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.005819940008223057, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.004848732613027096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.003472526790574193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.003460289677605033, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0033412035554647446, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0029914246406406164, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.002391578396782279, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0022366009652614594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0018720715306699276, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0017320560291409492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.001692466321401298, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.001639467547647655, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.001576092792674899, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0013224194990471005, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.001593312481418252, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0011939707910642028, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.001479910220950842, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.010789825581014156, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.052314527332782745, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.04943885654211044, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.04857482761144638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.04445774480700493, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.02284935861825943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.022046014666557312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.025143226608633995, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.023347247391939163, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.023066680878400803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.021109312772750854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.020400553941726685, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.012711968272924423, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.011197316460311413, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.010993901640176773, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.01094423234462738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.006396540440618992, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.006020486354827881, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.006007532589137554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.005719262640923262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.005710531957447529, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.0036074467934668064, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.004126736894249916, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.00354739255271852, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.003309252206236124, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.025143226608633995, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.061872951686382294, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.058509644120931625, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.05749166011810303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.05266064032912254, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.02698768675327301, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.02603728696703911, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.029671553522348404, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.027549654245376587, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0272354893386364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.024909671396017075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.024021124467253685, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.014869017526507378, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.012980375438928604, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.012742636725306511, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.012687929905951023, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.007408988196402788, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.006596153136342764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.006578507833182812, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.006208787672221661, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.006198007147759199, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.003923457115888596, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.004044819623231888, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.0038451198488473892, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.002796474378556013, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.029671553522348404, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.0.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.044239792972803116, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.03816048800945282, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.0352894589304924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.03208187222480774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.01897048018872738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.016604144126176834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.024110864847898483, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.021870644763112068, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.01985720358788967, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.016532810404896736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.016084952279925346, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.012246173806488514, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.010681742802262306, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.009418109431862831, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.009098896756768227, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.006380747072398663, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.005531215108931065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.00540988938882947, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.005138838198035955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.004972272552549839, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.0038969293236732483, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.0043047755025327206, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.0035472926683723927, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.0036342244129627943, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.03208187222480774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.011780252680182457, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.008928975090384483, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0049223024398088455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.004856664687395096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.004676425363868475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0020403244998306036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.009926741942763329, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.008504985831677914, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0054634492844343185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.00414355518296361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.004554252605885267, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.005026109982281923, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.004075726494193077, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.00241375295445323, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0018220666097477078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0025468512903898954, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0014848357532173395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0012570664985105395, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0014043061528354883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0011090944753959775, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.001413524616509676, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0013949500862509012, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0008697605226188898, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0010600258829072118, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.011780252680182457, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.01220211386680603, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.009076680988073349, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0050568715669214725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.004966003820300102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0047543831169605255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.002048776252195239, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.00963747687637806, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.008586762472987175, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.005591194145381451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.004177517723292112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.004508310463279486, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.004836062900722027, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.00409944262355566, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0024088069330900908, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.001798379817046225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.002446295227855444, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0014189506182447076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0011700541945174336, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0013269218616187572, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0009955849964171648, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0013147888239473104, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0013154663611203432, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.000750652514398098, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.000934409792535007, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.01220211386680603, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.08550907671451569, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06506945192813873, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05451371148228645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04639105498790741, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03631819412112236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02659127674996853, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.051224835216999054, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.04624940827488899, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03988247364759445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027398204430937767, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026881607249379158, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.026304949074983597, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.022242777049541473, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01765207201242447, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01640918292105198, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.013168229721486568, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.009456324391067028, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008975379168987274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007785331923514605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006971684750169516, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.006903087254613638, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.006916909944266081, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0052284784615039825, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004735325463116169, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027398204430937767, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08541550487279892, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06177406758069992, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.051301099359989166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04176291078329086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.036582101136446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.026654446497559547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04991908371448517, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.044519759714603424, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03959820419549942, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.025042366236448288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02470408007502556, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.025559280067682266, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02198415994644165, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.018537763506174088, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01764664240181446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.013270137831568718, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.011262128129601479, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010988698340952396, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00956039223819971, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009069343097507954, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.008061341010034084, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009178344160318375, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0070933811366558075, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007853907532989979, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.025042366236448288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.07473144680261612, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.07124627381563187, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.07027475535869598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.06475453078746796, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.03245965391397476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.031540967524051666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.03552727401256561, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.03291408345103264, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.032697003334760666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.03025216981768608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.029400905594229698, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.017909806221723557, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.015710096806287766, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.015520172193646431, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.015476485714316368, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.008995339274406433, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.008402587845921516, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.00839132722467184, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.008014746010303497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.008015956729650497, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.005029785912483931, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.005638272501528263, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.004971533082425594, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.004443889483809471, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.03291408345103264, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.0966516062617302, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.09233920276165009, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.09113277494907379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.08416521549224854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.04203812777996063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.04087945446372032, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.046039775013923645, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.042614731937646866, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.042345013469457626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.03924219682812691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.03813716024160385, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.023050660267472267, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.0200461745262146, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.019806353375315666, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.01975213550031185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.011479675769805908, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.010190613567829132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.010175601579248905, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.009663299657404423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.009664508514106274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.006086195819079876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.006143670063465834, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.00600765785202384, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.004160812124609947, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.023050660267472267, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.1.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.08005181699991226, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.07293523848056793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.06704707443714142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.06108567863702774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.035161472856998444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.03188604488968849, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.05050386115908623, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.0416400246322155, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.03604639321565628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.03148098289966583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.030676595866680145, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.02315530739724636, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.01952178031206131, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.01734315976500511, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.01695726439356804, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.012020670808851719, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.010197232477366924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.010409988462924957, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.009593935683369637, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.009799974970519543, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.00827784463763237, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.00783870555460453, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.00888768769800663, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.007252181880176067, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.03148098289966583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.014117065817117691, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.012061932124197483, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.010779693722724915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.009288785979151726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.006249005440622568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.005193485412746668, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.009060058742761612, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0076329829171299934, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.006525050383061171, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.005141944158822298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.004944228567183018, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.004636808764189482, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.003664052579551935, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0030467358883470297, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.002884629648178816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.002343610394746065, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.001671472447924316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0015980773605406284, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0014717687154188752, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0013780711451545358, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0012754845665767789, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0012286260025575757, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0009981503244489431, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0009120490285567939, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.014117065817117691, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.014021724462509155, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.011853896081447601, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.01058642566204071, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.009079962968826294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0060957870446145535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.005034209229052067, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.008104188367724419, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.007452109828591347, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.006377356592565775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.004977940581738949, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.004735519178211689, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.004088637884706259, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.003548484994098544, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.002943249884992838, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.002785483608022332, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.002065983135253191, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0015688266139477491, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0014976271195337176, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.001358266919851303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0012616179883480072, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.001116803614422679, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00110953813418746, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0009221722721122205, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0007660763221792877, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.014021724462509155, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.11043213307857513, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.09899018704891205, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.09522751718759537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.08336066454648972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.04987817257642746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.04565815255045891, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.05720554292201996, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.052705179899930954, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.05109791457653046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.042502790689468384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.03966578096151352, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.028951337561011314, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.025090791285037994, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.02380610629916191, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.02348010428249836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.014450140297412872, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.012105697765946388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.011999150738120079, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.010587465018033981, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.010458748787641525, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.007414986379444599, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.007233126554638147, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.006922596134245396, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0044473037123680115, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.028951337561011314, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1477786898612976, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12716229259967804, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12027275562286377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10218114405870438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06632669270038605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05838456377387047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07789487391710281, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07148416340351105, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06893490254878998, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.054040223360061646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04944383352994919, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04009772092103958, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03484800457954407, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.032561108469963074, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03200609236955643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02047722600400448, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0182090662419796, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01805599220097065, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01604750007390976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01583232544362545, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01196442823857069, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013102217577397823, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011295484378933907, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01063821092247963, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.032561108469963074, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.11790794879198074, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.11234963685274124, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.11078409105539322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.10220037400722504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.052310071885585785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.050714343786239624, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.05732552707195282, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.05307159945368767, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.05271249637007713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.04858045279979706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.047179751098155975, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.02882387675344944, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.025111455470323563, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.02477804385125637, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.024701615795493126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.014382448047399521, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01286518108099699, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.012850672006607056, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.012166565284132957, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.012166254222393036, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.0076830461621284485, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.00793528277426958, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.0075783138163387775, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.005577432457357645, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.02882387675344944, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1403263360261917, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.13383592665195465, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.13202054798603058, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.12193891406059265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.06237151473760605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.060478322207927704, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.06834643334150314, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.06326727569103241, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.06285595148801804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.05795334279537201, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.056291352957487106, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.034299060702323914, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.02982560731470585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.029438838362693787, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.02934667095541954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.017066676169633865, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.014971266500651836, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.014946920797228813, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.014108913019299507, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.014107579365372658, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.008895964361727238, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.008750563487410545, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.00877083744853735, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0055711958557367325, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.02982560731470585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.2.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.14194640517234802, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.13177762925624847, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.12786760926246643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.11761180311441422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.06367871165275574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.06018254533410072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.07326488196849823, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.06694354116916656, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.06482183933258057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.05793394893407822, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.056220751255750656, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.03686412051320076, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.031884606927633286, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.030453357845544815, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.030116071924567223, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.018533583730459213, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.016024988144636154, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.015874093398451805, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.014933095313608646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.014730249531567097, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.01018653903156519, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01031337771564722, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.00974612683057785, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.007434046361595392, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.031884606927633286, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0192439965903759, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.017107641324400902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.015846550464630127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.01387301180511713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0087313586845994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.007655363995581865, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.011147831566631794, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.010103658773005009, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.008987759239971638, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.007411076687276363, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.007044677156955004, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0056748283095657825, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0048440792597830296, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.004230716731399298, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.004075446166098118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0028667435981333256, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.002268280601128936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0021970239467918873, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.002018765779212117, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0019326047040522099, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00156397873070091, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.001574800000526011, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0013519002823159099, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0011374386958777905, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0192439965903759, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.019733956083655357, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.017438843846321106, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.015949729830026627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.013890515081584454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.008804356679320335, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0076011731289327145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.011304829269647598, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.010495083406567574, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.009078589268028736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.007453124038875103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.007086026482284069, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.005708516109734774, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.004981962963938713, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.00424726027995348, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.004056818317621946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.002868517069146037, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.002241348149254918, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0021478126291185617, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.001979677937924862, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0018677293555811048, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.001528857508674264, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.001539477612823248, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0012956620194017887, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0010471565183252096, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.019733956083655357, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.12397593259811401, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.11285020411014557, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.10917411744594574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0967421904206276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.05685000494122505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05265451967716217, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0649452731013298, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.05958274379372597, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.058056335896253586, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.04920351505279541, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.04619552940130234, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03290349990129471, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.02836633287370205, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.027122756466269493, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.026823999360203743, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.016427848488092422, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.013754038140177727, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.013654724694788456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01217274647206068, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.012062685564160347, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.008401427417993546, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008097690530121326, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.007888996973633766, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.004890157841145992, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03290349990129471, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14475974440574646, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12218005210161209, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11403042078018188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09314167499542236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06499135494232178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0565672367811203, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07860428839921951, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0717615857720375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0679490715265274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05121467635035515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04593968018889427, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04052570089697838, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.035334356129169464, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03232227638363838, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03159625828266144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020860156044363976, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018662795424461365, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018443334847688675, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01636805757880211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01604200340807438, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012466548942029476, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014090762473642826, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011622396297752857, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011805475689470768, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03232227638363838, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.14778800308704376, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.14036889374256134, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.13825927674770355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.12725001573562622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.06743171066045761, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.06509606540203094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.07428837567567825, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.06860929727554321, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0680307149887085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.062189020216464996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06027666851878166, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.03792613372206688, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03324158117175102, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03272894397377968, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.032607756555080414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.01908206194639206, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01800556853413582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.017974799498915672, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01707274839282036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01705937460064888, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.010832058265805244, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01240704208612442, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010671873576939106, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.010037424974143505, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03324158117175102, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1729646474123001, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16433025896549225, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16183702647686005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14900071918964386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07859748601913452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0758606567978859, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08654867112636566, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08000031858682632, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07933761179447174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07242604345083237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07008757442235947, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.043817367404699326, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03806198760867119, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03745414689183235, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.037311021238565445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.021867243573069572, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.019488707184791565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.019452307373285294, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018301356583833694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018291784450411797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011704076081514359, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012068122625350952, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01150903757661581, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00850511435419321, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.021867243573069572, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.3.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.07770571857690811, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.07261350750923157, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.05094704404473305, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.04689791426062584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.03695237636566162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.025191770866513252, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.07490961253643036, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.06197822466492653, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.0374157652258873, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.034385208040475845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.036173831671476364, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.03652884066104889, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.030679015442728996, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.01969054713845253, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.01785367541015148, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.020008578896522522, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.015080510638654232, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.014246013946831226, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.014785044826567173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.013896857388317585, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015964748337864876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01435674075037241, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014623450115323067, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.013215543702244759, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.025191770866513252, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.02554512955248356, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.023014109581708908, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.021707553416490555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.019232919439673424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.011634896509349346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.010463373735547066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.014083691872656345, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.012990829534828663, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.011925977654755116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.010025177150964737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.009521638974547386, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.007145935203880072, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.006214286666363478, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.00561258289963007, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.005462816916406155, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0035910068545490503, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0029702060855925083, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.002903454005718231, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0026601417921483517, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.002581141423434019, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0019233535276725888, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.001982372486963868, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0017289073439314961, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.001404264708980918, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.02554512955248356, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.026920486241579056, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.02413046732544899, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.02248391881585121, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.019821282476186752, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.01206013560295105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.010681173764169216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.014967133291065693, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.013872971758246422, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.012383155524730682, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.010368268936872482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.00986451841890812, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.007549687288701534, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.00660735135897994, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.005803778767585754, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.005595536902546883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.003787981579080224, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.003038282971829176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0029400361236184835, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.002707851119339466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0025917701423168182, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.002014740603044629, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.002025546273216605, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.001758531667292118, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0013680103002116084, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.026920486241579056, + "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": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13482294976711273, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.12399104237556458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.12047303467988968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.10768657177686691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06191718205809593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0578949898481369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07023011893033981, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06435711681842804, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06300647556781769, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05429338663816452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.051311276853084564, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03558136150240898, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030636277049779892, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.029515059664845467, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.029242102056741714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01772977039217949, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.014949819073081017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.014859812334179878, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013381714932620525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.013290783390402794, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.009074514731764793, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.008757275529205799, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008593556471168995, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00529692554846406, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.030636277049779892, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1435447633266449, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12532824277877808, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11675538122653961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09887364506721497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06532878428697586, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05721640959382057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08135416358709335, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0737968236207962, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06737912446260452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053602706640958786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.050249505788087845, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04172244295477867, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0359819121658802, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03221910446882248, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03128344565629959, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02120307832956314, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018196292221546173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01781470514833927, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01615791581571102, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015701932832598686, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01221169251948595, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013470564968883991, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011055623181164265, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010898029431700706, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03221910446882248, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.13994236290454865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.13239893317222595, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.13018669188022614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.11942248046398163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.06391393393278122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.061440691351890564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.0703781247138977, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.06514593213796616, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.06455644965171814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.058463204652071, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.05642249062657356, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.035785600543022156, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.031165452674031258, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03062324970960617, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.030503472313284874, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.017901308834552765, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01616021990776062, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.01612451858818531, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.015124511905014515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.015114234760403633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.009709484875202179, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.010329569689929485, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.009541296400129795, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007619088049978018, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.031165452674031258, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17283529043197632, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16360582411289215, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16091232001781464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.147628054022789, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07901456207036972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07595833390951157, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08701299875974655, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08051697164773941, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07980241626501083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07225440442562103, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06967098265886307, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.044051676988601685, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.038339924067258835, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03767596185207367, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03752756863832474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02198176272213459, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01956956647336483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.019525030627846718, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01825074478983879, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018240224570035934, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011671114712953568, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012063560076057911, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011457293294370174, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00842341035604477, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02198176272213459, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.4.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.05785040184855461, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.05174870789051056, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.047915179282426834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.04259524866938591, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.026492103934288025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.023639146238565445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.03612641245126724, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.03069513477385044, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.02719392627477646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.022854264825582504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.022160176187753677, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.017459852620959282, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.014583731070160866, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.012983301654458046, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.012519022449851036, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.008636072278022766, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.007054909132421017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.006871046964079142, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.006379918660968542, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.006147344596683979, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.005156402010470629, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.004989433102309704, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.004504871089011431, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.0038117244839668274, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.03069513477385044, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0378710962831974, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.033857204020023346, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.032104045152664185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.02833271585404873, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.017170928418636322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.015476534143090248, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.02033797837793827, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.018802177160978317, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.017638573423027992, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.014659267850220203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.013847019523382187, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.010296913795173168, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.008969645015895367, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.008237568661570549, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.008060617372393608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.005154827609658241, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0042958431877195835, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0042221397161483765, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.003803154220804572, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0037122543435543776, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0027183163911104202, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.002765584271401167, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0024776558857411146, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0018825976876541972, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.032104045152664185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.039560362696647644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.03511371463537216, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03313792496919632, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.029117494821548462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.017690036445856094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.01580517180263996, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.021295832470059395, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.019628938287496567, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.01821906864643097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.015050847083330154, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.014256646856665611, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01074141450226307, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.009328952059149742, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.00847263727337122, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.008260770700871944, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.005373324733227491, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004379682708531618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0042877765372395515, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.003855982329696417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0037407793570309877, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0028043570928275585, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0027954380493611097, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.002515337662771344, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0018270299769937992, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03313792496919632, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.13918787240982056, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1280907541513443, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1245269924402237, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.11132314801216125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.06416084617376328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.05999186635017395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.07229092717170715, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.06669674813747406, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.06530121713876724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.05626772716641426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.05315317586064339, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03660346567630768, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03177782520651817, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.030618660151958466, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.030347615480422974, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01825586147606373, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01554329413920641, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.015448804944753647, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.013916771858930588, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.013831517659127712, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00936932023614645, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.009136957116425037, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.008935902267694473, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005582165904343128, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03177782520651817, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14190827310085297, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1212797686457634, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11349022388458252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09304208308458328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06508950144052505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05655527114868164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0773102194070816, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0709066092967987, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06707771122455597, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.050618063658475876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04601607844233513, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03991800546646118, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.034407202154397964, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03175457939505577, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03110477700829506, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.020106391981244087, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01730566844344139, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0170780997723341, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014662545174360275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01437927596271038, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.011204740963876247, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012051188386976719, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.010365158319473267, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009316563606262207, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03175457939505577, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.15469275414943695, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.14604279398918152, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.14350809156894684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1310606598854065, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07103919237852097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.068092942237854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.07833681255578995, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07248232513666153, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07179383188486099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0645299181342125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06198939308524132, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.039689045399427414, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03450809419155121, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03386523202061653, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03371132165193558, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.019817501306533813, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01748770661652088, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.017447587102651596, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.016200587153434753, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01618361659348011, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.010494960471987724, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.0106358015909791, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010284225456416607, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007238571532070637, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.019817501306533813, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.19143414497375488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.18082007765769958, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1777079552412033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.16230668127536774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0879334956407547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0843096524477005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09689018875360489, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08971243351697922, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08889298886060715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07991708070039749, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07670369744300842, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04897448420524597, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04261503741145134, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04182260483503342, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04163236916065216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024409761652350426, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021285004913806915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02123448997735977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019655201584100723, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019640851765871048, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012663540430366993, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01247294433414936, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01240248791873455, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007879728451371193, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024409761652350426, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.5.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.18921828269958496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1718565821647644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.16548998653888702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.14776550233364105, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.08582250028848648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.0797860249876976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.09879652410745621, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.09077348560094833, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.08767858147621155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07516670972108841, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07101291418075562, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05000089108943939, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04334670677781105, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.041154030710458755, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04062807187438011, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.025066357105970383, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.021654460579156876, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.021424755454063416, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.019621696323156357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.01929454132914543, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.013594143092632294, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.013963627628982067, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.012918056920170784, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010013056918978691, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.025066357105970383, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04282473400235176, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0387340784072876, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.03686809539794922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03277220204472542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.019554613158106804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.017778856679797173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.023118585348129272, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.021375330165028572, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02003416419029236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.01690606214106083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.016031187027692795, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01170828752219677, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.010205206461250782, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.009397666901350021, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.00920010544359684, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.005861076060682535, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0049132383428514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.004829746671020985, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.004392058122903109, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.004292431753128767, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.003092625178396702, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.003175410907715559, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00282393884845078, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.002186079975217581, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03277220204472542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.041520874947309494, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.03726533055305481, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.03507246822118759, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.031055554747581482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.01869492046535015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.016751758754253387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.022746697068214417, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.020977335050702095, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.019189199432730675, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.016117654740810394, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.015337402932345867, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.011483498848974705, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.00998822133988142, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.008967560715973377, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.008714893832802773, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.005749328061938286, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004646599758416414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.004530787002295256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004136078990995884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.003998148255050182, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0030143619515001774, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0030009434558451176, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0026799156330525875, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0019702091813087463, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.031055554747581482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1517959088087082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.14019662141799927, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.13643650710582733, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.12248854339122772, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0700409859418869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0657133087515831, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0788942351937294, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.07279282808303833, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.07127264887094498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0617181621491909, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.058472782373428345, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.03992787003517151, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03466308116912842, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.033421922475099564, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03311825171113014, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01991633139550686, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0169150922447443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01682111993432045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.015191465616226196, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.015094668604433537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.010188048705458641, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00987390149384737, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.009723843075335026, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.005929499864578247, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.033421922475099564, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1701006442308426, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14158742129802704, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.13179980218410492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11110605299472809, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07719901949167252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06560521572828293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09105964750051498, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08332741260528564, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08005549758672714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05949179083108902, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05468438193202019, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04697278141975403, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.040512412786483765, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03777144476771355, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03708914667367935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023926803842186928, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0207709688693285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020574145019054413, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01758890226483345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017275400459766388, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013843289576470852, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01460479386150837, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013053258880972862, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011518226005136967, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023926803842186928, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16185195744037628, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1527068018913269, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15000727772712708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.13702738285064697, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07465692609548569, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07148761302232742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08230370283126831, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0762212872505188, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07548356801271439, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06774935126304626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06506282836198807, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.041712839156389236, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.036315180361270905, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03561966121196747, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.035455573350191116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.020817440003156662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.018379729241132736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.018330037593841553, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.017005931586027145, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.0169847272336483, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01097018364816904, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01115360390394926, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010741397738456726, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0075530135072767735, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.020817440003156662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.19856035709381104, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.18742425739765167, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1841164082288742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.16823077201843262, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.09157189726829529, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.087717704474926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10084962844848633, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09346019476652145, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0925864726305008, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08311741054058075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07973159104585648, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05100386589765549, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.044424548745155334, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04357381537556648, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.0433785542845726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.025416625663638115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.022148994728922844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.022096965461969376, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02042710781097412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.020402994006872177, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.013118821196258068, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012932832352817059, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01283908262848854, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008092219941318035, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.025416625663638115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.6.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.1999904364347458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18160010874271393, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17486877739429474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15586742758750916, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09090360999107361, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08445359766483307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10450415313243866, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.09609697759151459, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09285327047109604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07944013178348541, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07492157816886902, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05301256477832794, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04584086313843727, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04355710744857788, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04301206022500992, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.026545239612460136, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02281840704381466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02257411926984787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.020610088482499123, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02026408724486828, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.014352545142173767, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.014576778747141361, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.013625046238303185, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010304881259799004, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.026545239612460136, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.04136987403035164, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.038035567849874496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.03648926317691803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03272821754217148, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.01910206861793995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.017618760466575623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.022301336750388145, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.020616618916392326, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.01945618912577629, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.016776036471128464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.01597096398472786, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.011311381123960018, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.009860592894256115, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.00918666273355484, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.00902244821190834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.005666074343025684, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.004809608682990074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.004737300798296928, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.004352656193077564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.004272946156561375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0029998142272233963, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.003095559310168028, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0027793985791504383, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0021538743749260902, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.03272821754217148, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.03723645955324173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0341557078063488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.032324425876140594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.028951574116945267, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.017012491822242737, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.015471438877284527, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.020598335191607475, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.019026901572942734, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.017345499247312546, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.014947757124900818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.014311953447759151, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01039238553494215, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.009057268500328064, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.008158288896083832, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.007934327237308025, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.005201826337724924, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.004227765835821629, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00412205932661891, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.003816383657976985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.003699279623106122, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0027359682135283947, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0027195068541914225, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0024541765451431274, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0017905151471495628, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.032324425876140594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.16596214473247528, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.15407684445381165, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.15040205419063568, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1354762315750122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.07677461951971054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07245707511901855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0865134745836258, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.07935276627540588, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.07795482873916626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0680808573961258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.06470862776041031, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04382063448429108, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.037798717617988586, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.036628734320402145, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03635505959391594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0218658447265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.018525220453739166, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.018433645367622375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01672150008380413, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.016642024740576744, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.011177001520991325, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.010742837563157082, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.010679055005311966, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006421787664294243, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0218658447265625, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15662133693695068, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13142645359039307, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12036114931106567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09839306771755219, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07074553519487381, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05966147035360336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08726273477077484, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08042203634977341, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07368285208940506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05491244047880173, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04954661801457405, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04490023851394653, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03894813731312752, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.034583695232868195, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03348085656762123, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022609759122133255, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018993325531482697, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01857381872832775, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016147324815392494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015582935884594917, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012593824416399002, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013597523793578148, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011272846721112728, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010487835854291916, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03348085656762123, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16058054566383362, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.15156857669353485, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.14891685545444489, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.13620026409626007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07432028651237488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0712040364742279, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08204515278339386, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07587121427059174, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07514744997024536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06749892234802246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06492619216442108, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.041687972843647, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03623596578836441, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03554784506559372, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03538491204380989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02082842029631138, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.018455956131219864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.018413394689559937, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.017110280692577362, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.017094803974032402, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011092373169958591, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011365711688995361, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010872145183384418, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007901778444647789, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02082842029631138, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.20269504189491272, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.19143445789813995, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.18811720609664917, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.17206498980522156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0937969833612442, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0898565724492073, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10339363664388657, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0957246869802475, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.09481427073478699, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08517301082611084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.0818401649594307, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05235113203525543, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04553800821304321, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04467254877090454, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04446585103869438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02609529159963131, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.022728390991687775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.022670917212963104, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.020974626764655113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.020953526720404625, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.013522318564355373, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.013305127620697021, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01323418878018856, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008375733159482479, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02609529159963131, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.7.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.20173916220664978, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1829979568719864, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17618249356746674, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15677344799041748, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09174405038356781, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08518768846988678, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10554195940494537, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.09694068133831024, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09373175352811813, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07996590435504913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07539166510105133, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.053553324192762375, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.046293795108795166, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04399823024868965, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.043448738753795624, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02687002904713154, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023101618513464928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02286183275282383, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.020844779908657074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.020497126504778862, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.01464769896119833, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.014824516139924526, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01393770519644022, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010564928874373436, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02687002904713154, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05498187243938446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05080211162567139, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.04908891022205353, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.044144634157419205, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02538965456187725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.023666825145483017, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.029084190726280212, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.026905469596385956, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.025827059522271156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.02242359146475792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02133977971971035, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.014759275130927563, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.012871772050857544, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.012189160101115704, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.012025346979498863, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.007385943550616503, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.006356947124004364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.006290241610258818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.005772655364125967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.005701630841940641, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0038973079062998295, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004027041140943766, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.003668451216071844, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0027894724626094103, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.029084190726280212, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04914944991469383, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.045310597866773605, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04341735690832138, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.038933586329221725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.022434134036302567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.020715100690722466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.026350131258368492, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02435206063091755, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.022843701764941216, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.019802208989858627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.018924465402960777, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.013306628912687302, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.011612538248300552, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.010749242268502712, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.010536949150264263, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.006658432073891163, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0055480957962572575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00545590091496706, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005025586578994989, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00491799833253026, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00348396273329854, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0034883718471974134, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003209219081327319, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0022983658127486706, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.026350131258368492, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17450156807899475, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.16269256174564362, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.15900646150112152, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.14369580149650574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08075162768363953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0764174535870552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09047485888004303, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08330248296260834, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08192528784275055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07201451063156128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.06850414723157883, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.04580862447619438, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.03963717073202133, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.038489893078804016, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.038217801600694656, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.022840335965156555, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.019464990124106407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.019373351708054543, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01765552908182144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.017583806067705154, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.011687198653817177, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.011264711618423462, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.011220358312129974, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006749051623046398, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.022840335965156555, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18431079387664795, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16083085536956787, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.153021901845932, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1279398649930954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08345670998096466, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07459899038076401, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.096893809735775, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08925125002861023, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08630486577749252, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0673208236694336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06167404353618622, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.049582503736019135, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.042916521430015564, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04031423479318619, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03967990353703499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02488059364259243, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02144552208483219, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021268747746944427, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018380658701062202, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01811988651752472, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013590455986559391, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014207514934241772, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012778223492205143, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01046344917267561, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02488059364259243, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1664244383573532, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1569797247648239, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15418671071529388, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14087100327014923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07707836478948593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07374726235866547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08508378267288208, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07872281223535538, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07793101668357849, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0698762834072113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06710021942853928, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04312173277139664, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03749031201004982, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03675268217921257, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.0365772545337677, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.021524306386709213, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01883557252585888, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.018788333982229233, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.017384037375450134, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01736098900437355, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01128268614411354, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011241456493735313, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01103949174284935, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0073592038825154305, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.021524306386709213, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2103062868118286, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.19838693737983704, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.19490523636341095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.17805549502372742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.09748811274766922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09332280606031418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10751255601644516, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09956265985965729, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.09858698397874832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08839032799005508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08481353521347046, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05445536598563194, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04737863689661026, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04645047336816788, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04622935876250267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02714879810810089, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.023608777672052383, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.023547323420643806, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.021746529266238213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.021717583760619164, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01404659915715456, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.013786901719868183, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.013736527413129807, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008615069091320038, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02714879810810089, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.8.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.20829731225967407, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18879754841327667, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.1816771775484085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16154439747333527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09477363526821136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08791211992502213, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10879060626029968, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10016216337680817, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09685887396335602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08243510127067566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07763270288705826, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05518607422709465, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.047779083251953125, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04541926085948944, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.044852107763290405, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02763589657843113, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023769302293658257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02352656051516533, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021393194794654846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021035313606262207, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.014938576146960258, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015153131447732449, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014204458333551884, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010682100430130959, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02763589657843113, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.050981178879737854, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.047372255474328995, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.04582080617547035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04134824499487877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02362780272960663, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.022089961916208267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.027107499539852142, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.02507825940847397, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.023994388058781624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.020993387326598167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02004009298980236, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.013753039762377739, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01198277622461319, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.011335153132677078, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.011179478839039803, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.0068825287744402885, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00589045649394393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.005823667161166668, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.005366933532059193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.005298524163663387, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0036199423484504223, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0037043073680251837, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.003402737434953451, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.002528988290578127, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.027107499539852142, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.04638488218188286, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.042982976883649826, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.04126463085412979, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.037140022963285446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.021272223442792892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.019702479243278503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.024892110377550125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.023072440177202225, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.021620338782668114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.018874116241931915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.01806041970849037, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.012574214488267899, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.011000819504261017, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.010195888578891754, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.009994564577937126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.006295893806964159, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00527260871604085, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0051800948567688465, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.004794016480445862, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.004699665121734142, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0033057306427508593, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003325031604617834, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0030576717108488083, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0022062554489821196, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.024892110377550125, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1832253485918045, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1711106300354004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.16745568811893463, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1514563262462616, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08491240441799164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08043834567070007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0950918048620224, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08715549111366272, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08599246293306351, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07574234902858734, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07222483307123184, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.048133060336112976, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04154417663812637, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.040466535836458206, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04020488262176514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02400650456547737, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.020480088889598846, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.020393934100866318, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.018596133217215538, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.018539581447839737, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.012296400032937527, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.011856873519718647, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01181777287274599, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007154173217713833, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02400650456547737, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17771577835083008, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1517249196767807, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14303448796272278, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12064249813556671, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08029524236917496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0704270526766777, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0940743163228035, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0866454541683197, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08344675600528717, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06404513120651245, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.057966236025094986, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04823068156838417, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04179980978369713, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.038943372666835785, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.038252148777246475, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024265306070446968, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02093529887497425, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02074015699326992, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01789582148194313, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017574213445186615, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013382603414356709, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014176041819155216, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012504099868237972, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010674728080630302, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024265306070446968, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16149219870567322, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.15241476893424988, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.14971134066581726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.13685692846775055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0749792531132698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07175811380147934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08282968401908875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07660233229398727, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07582694292068481, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0680074468255043, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06538338959217072, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04206617549061775, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03654249757528305, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03581402078270912, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03565094247460365, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02104261703789234, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.018476340919733047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.018430253490805626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01708599179983139, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.017061585560441017, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011151408776640892, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01120553258806467, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010912916623055935, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0075696599669754505, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02104261703789234, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2076629400253296, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1959718018770218, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.19251534342765808, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.176045224070549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.09645421802997589, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09232407063245773, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10639585554599762, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09848982840776443, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.09754428267478943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08750426769256592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08399982005357742, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05396484211087227, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.046910595148801804, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04599221423268318, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04577499255537987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02691144496202469, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.023430228233337402, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02337011694908142, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.021602129563689232, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02157224714756012, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.013988524675369263, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01376391015946865, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.013685480691492558, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00871312990784645, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02691144496202469, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.9.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2050560861825943, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18607524037361145, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17915378510951996, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15928269922733307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09331624954938889, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08668075501918793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10727742314338684, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.0986856073141098, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09538488835096359, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08139777928590775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.076565220952034, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05444559082388878, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.047226764261722565, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04485262185335159, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04428467899560928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027327746152877808, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02370857261121273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023466795682907104, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021445993334054947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021093666553497314, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.014901570975780487, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015432130545377731, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01417103223502636, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011226850561797619, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027327746152877808, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06349971890449524, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0589846596121788, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.057228341698646545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05163848400115967, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.029367024078965187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.027547771111130714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03337854892015457, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.030863888561725616, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.029822707176208496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.026080790907144547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.024869367480278015, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.016936149448156357, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.014750074595212936, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014082447625696659, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.013924147933721542, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.008472215384244919, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007315424736589193, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007252780720591545, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006664101500064135, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006601491943001747, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004460838157683611, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004576279781758785, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004235429689288139, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0031398353166878223, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03337854892015457, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05426536500453949, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05028875172138214, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.048369377851486206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.043502986431121826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.024820007383823395, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.023050900548696518, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02887747623026371, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.026713263243436813, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.025206467136740685, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.022014517337083817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.021026013419032097, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.014558578841388226, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.012722061946988106, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01187166292220354, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011665742844343185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007287210319191217, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006118034012615681, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00602411525323987, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005554836709052324, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005457872059196234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0038146055303514004, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003809605725109577, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003548351814970374, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0024991596583276987, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02887747623026371, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.17727944254875183, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.16563154757022858, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.16197313368320465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1466359943151474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08208980411291122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07785315066576004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09193681925535202, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08455781638622284, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08325142413377762, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07338833808898926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07003439962863922, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.046577151864767075, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04025783762335777, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.03914150223135948, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0388733446598053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.023221885785460472, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.01979859732091427, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.01971789076924324, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.017997434362769127, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.017927465960383415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01187726017087698, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01147013995796442, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.011416536755859852, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.006904226262122393, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.023221885785460472, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17017285525798798, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1489204466342926, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14140719175338745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12144499272108078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07723648101091385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06902819871902466, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09089381247758865, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08319980651140213, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07989957928657532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06347185373306274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05936457961797714, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.046878382563591, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.040360018610954285, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.037681929767131805, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03703681007027626, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023788003250956535, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02058824710547924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.020390428602695465, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018059363588690758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017773665487766266, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013612736016511917, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014300446026027203, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012823455035686493, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01114699523895979, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.023788003250956535, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1583070009946823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.14932380616664886, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1466609388589859, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1340392380952835, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07355225086212158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07039407640695572, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08119558542966843, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07516590505838394, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07438862323760986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06668688356876373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06406006962060928, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04127000272274017, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03587448596954346, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03516094386577606, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03498835116624832, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02059914544224739, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.018147170543670654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.01810140162706375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.016773931682109833, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.016751248389482498, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.010872538201510906, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011022445745766163, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010639717802405357, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007464686408638954, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02059914544224739, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.20860692858695984, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.19681784510612488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.19332335889339447, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.17674794793128967, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.09694728255271912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0927729606628418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10694121569395065, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09902209043502808, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0980541855096817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08788351714611053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08433296531438828, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.0542028546333313, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.047151364386081696, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.046219244599342346, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04599794000387192, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.027023572474718094, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02353503182530403, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.023475611582398415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02168465591967106, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.021655883640050888, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014002096839249134, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.013807152397930622, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01368976105004549, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008716566488146782, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.027023572474718094, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.10.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.20289486646652222, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1838596761226654, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17698870599269867, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.1575026959180832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.0923137441277504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08564893901348114, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10639884322881699, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.09764958173036575, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09438007324934006, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.0804673358798027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07574775069952011, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.053916264325380325, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04674512520432472, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04439413547515869, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04382520914077759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02710879035294056, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023490184918045998, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023243511095643044, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021238552406430244, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.020883573219180107, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.014860196970403194, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015311090275645256, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014127449132502079, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.01115564163774252, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02710879035294056, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.05954040214419365, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05535200610756874, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.053628142923116684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04840566962957382, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02755282074213028, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02582210674881935, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03147745132446289, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.02911144495010376, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.027974572032690048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.024508317932486534, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.023388171568512917, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.015977151691913605, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.013915679417550564, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.013223858550190926, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.013058057054877281, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.007992899045348167, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.006887412630021572, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0068191708996891975, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.006286279764026403, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.006216229870915413, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004221627488732338, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004345342051237822, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.003992709796875715, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003002564888447523, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03147745132446289, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05213429406285286, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.04832986742258072, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.046359360218048096, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.041701849550008774, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.023804672062397003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.022062724456191063, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02792903035879135, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.02585022523999214, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02417144365608692, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.021141523495316505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.020246872678399086, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01411542110145092, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.012306643649935722, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.011394469067454338, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011174498125910759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007052650209516287, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0058799562975764275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00577532546594739, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005348526872694492, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.005235899239778519, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.003700635861605406, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0036888413596898317, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0034128455445170403, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0024211222771555185, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02792903035879135, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1861838549375534, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.17390450835227966, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.17017307877540588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.15380564332008362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08616094291210175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08170502632856369, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09655989706516266, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08866391330957413, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08735650032758713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07697635143995285, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07338006049394608, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0488964319229126, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.042197033762931824, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.041056327521800995, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04078959301114082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.024395659565925598, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.020784862339496613, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.020694874227046967, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01888500340282917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.018815606832504272, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.012486107647418976, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.012063302099704742, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012000457383692265, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007286793552339077, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.024395659565925598, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18442398309707642, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16037698090076447, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15180157124996185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12761060893535614, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08419375866651535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07458608597517014, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09832552075386047, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09056080877780914, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08667121082544327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06705216318368912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.061931949108839035, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.050396889448165894, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.043559323996305466, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.040672414004802704, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03997102379798889, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025246618315577507, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02156376838684082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021322892978787422, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01828235574066639, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017968056723475456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013726804405450821, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014242764562368393, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012814493849873543, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010360777378082275, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025246618315577507, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.14625735580921173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.13795310258865356, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.13546380400657654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.12384078651666641, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.06795097887516022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.06502051651477814, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.07500237971544266, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.06944543123245239, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.06873293220996857, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.061602212488651276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.05920808017253876, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.03812338039278984, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.033179085701704025, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.032511722296476364, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03235334903001785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.019042130559682846, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01685362495481968, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.016811374574899673, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01559427473694086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.015571502968668938, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.010073307901620865, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.010348279029130936, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.00985799077898264, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007140791974961758, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.033179085701704025, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2027331292629242, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.19128474593162537, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1878543496131897, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.17173287272453308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.09415937960147858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09009918570518494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10383908450603485, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09620258957147598, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.09522838145494461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08534827083349228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08191505819559097, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.052666496485471725, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04581015557050705, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04489560425281525, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04467995464801788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.026254532858729362, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.022881560027599335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02282087318599224, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.021082710474729538, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.021053621545433998, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.013624276965856552, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01345270499587059, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.013319285586476326, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00852629728615284, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.026254532858729362, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.11.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2033514380455017, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18449848890304565, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17758852243423462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15800081193447113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.0924522653222084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08590290695428848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10640040785074234, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.09780410677194595, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09450482577085495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08058100938796997, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.0759153664112091, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05387664958834648, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.046733323484659195, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04440748319029808, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04384222254157066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027095751836895943, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02342679351568222, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023180674761533737, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021158358082175255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02080680802464485, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.014827786944806576, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015176581218838692, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014100815169513226, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010965652763843536, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027095751836895943, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06260237097740173, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.058390505611896515, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0567394495010376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0513661690056324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.02902773581445217, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.027333185076713562, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03286438062787056, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.030448472127318382, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.02944890223443508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.025929557159543037, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.024763070046901703, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.016675148159265518, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.014541835524141788, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.013911225832998753, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.013762678019702435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.008338545449078083, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0072078523226082325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007147438824176788, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0065910243429243565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0065301209688186646, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0043855877593159676, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004475297871977091, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00417659617960453, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0030352973844856024, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03286438062787056, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05529763549566269, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.051468513906002045, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.049693893641233444, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0448312945663929, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02535063400864601, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.023685485124588013, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02929268777370453, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.027110131457448006, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.025716053321957588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.022623024880886078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02164691872894764, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01476655900478363, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01291242241859436, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01211613416671753, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.011925769038498402, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0073878769762814045, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006219062954187393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006130477879196405, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.005672674626111984, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0055798995308578014, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0038551962934434414, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.003825114807114005, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003609489882364869, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002463224111124873, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.02929268777370453, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.18785370886325836, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.17573235929012299, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.17206205427646637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1558723747730255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08698149025440216, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08265964686870575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09710264205932617, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08930070698261261, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08812230080366135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07787665724754333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07438389211893082, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.049159854650497437, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.042486660182476044, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04143413156270981, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.041187968105077744, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02450716868042946, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.020930109545588493, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.020853787660598755, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0190553218126297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.018999453634023666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.012527995742857456, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01206605788320303, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012081480585038662, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0072108907625079155, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02450716868042946, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19089575111865997, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16157297790050507, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15181441605091095, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12174151092767715, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0859738364815712, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07482117414474487, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10125473141670227, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09317413717508316, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08955751359462738, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06603633612394333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0584944412112236, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.051935840398073196, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04485539346933365, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.041604671627283096, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04081286862492561, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026103243231773376, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022214477881789207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021980924531817436, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018457716330885887, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018097246065735817, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014397799968719482, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014896554872393608, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013397132977843285, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010998964309692383, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026103243231773376, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1555889993906021, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1467105746269226, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.14404182136058807, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.13170099258422852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07248306274414062, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.06933197379112244, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08010385185480118, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07407067716121674, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07331611961126328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06566783040761948, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06317130476236343, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.040845029056072235, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03548695519566536, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03477175533771515, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03461120277643204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02041962742805481, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.018179897218942642, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.018131151795387268, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.016846412792801857, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.016818812116980553, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.010966935195028782, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011379141360521317, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010741963051259518, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008103443309664726, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02041962742805481, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2101968228816986, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.19822974503040314, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.19469721615314484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.17792578041553497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.09775404632091522, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09351892024278641, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10787276923656464, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09987640380859375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.09887765347957611, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08856187760829926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08497181534767151, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.0547323152422905, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04756993055343628, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04662120342254639, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04639490693807602, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02728957124054432, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.023746250197291374, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.023685317486524582, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.021859580650925636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02182944491505623, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014177187345921993, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.013939820230007172, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.013862313702702522, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008809699676930904, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02728957124054432, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.12.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2097151279449463, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19052423536777496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18345944583415985, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16319148242473602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09551434218883514, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08872577548027039, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10980725288391113, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10096684098243713, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09757831692695618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08326013386249542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07844910770654678, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.0557456910610199, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04823127016425133, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.045814454555511475, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04523949325084686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027921078726649284, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02406039647758007, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023805366829037666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02170281857252121, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02133459784090519, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015105520375072956, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015445212833583355, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014341169036924839, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010994094423949718, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027921078726649284, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06750740110874176, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06291425973176956, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06109035015106201, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05524072051048279, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03129727765917778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02942754328250885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03554758056998253, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03290028125047684, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03177144378423691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.027916215360164642, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.026669548824429512, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018028827384114265, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015712881460785866, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014995075762271881, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014822556637227535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009014669805765152, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007756682112812996, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0076872967183589935, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007079876959323883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007010228931903839, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0047303359024226665, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0048026531003415585, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004493107087910175, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0032313799019902945, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03290028125047684, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05980726704001427, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.055616095662117004, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05349721387028694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04825098440051079, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.027378063648939133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.025515194982290268, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.031842343509197235, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.029534339904785156, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.027822524309158325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.024391891434788704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02334047481417656, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01604059338569641, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014023549854755402, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013093715533614159, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012872759252786636, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008015591651201248, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006729131564497948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006624320987612009, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006134231109172106, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006022609770298004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00419349130243063, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004171003587543964, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0039035561494529247, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002697285730391741, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.031842343509197235, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.19666770100593567, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18425163626670837, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1803949475288391, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16352280974388123, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09140389412641525, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08681201934814453, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10218127816915512, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09386538714170456, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09261257946491241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08186136186122894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07815500348806381, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05174342915415764, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04470926895737648, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04357164353132248, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.043291281908750534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02580970898270607, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.022019101306796074, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.021937012672424316, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02004830539226532, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.01998276263475418, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013198058120906353, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.012710696086287498, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012723037973046303, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007601473946124315, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02580970898270607, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1912706047296524, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16629521548748016, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1579730063676834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1275608241558075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08672010153532028, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07717952877283096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1005149632692337, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0926937609910965, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08946368843317032, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06851198524236679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06130727380514145, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05132991448044777, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.044490914791822433, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04180488735437393, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04115266352891922, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025807417929172516, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022089751437306404, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02189001999795437, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018590344116091728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018356213346123695, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014066038653254509, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014436597004532814, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013237598352134228, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010413995943963528, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025807417929172516, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1598796844482422, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.15078838169574738, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.14806526899337769, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1353055089712143, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07451524585485458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07126990705728531, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08233988285064697, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07616303861141205, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0753667950630188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06749758869409561, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06488689035177231, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04190763458609581, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.036402277648448944, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03567076101899147, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.035498760640621185, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.020936081185936928, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.018486184999346733, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.01844145730137825, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01709197461605072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01706731505692005, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011110840365290642, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011342895217239857, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.010871878825128078, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007816859520971775, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.020936081185936928, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.21915048360824585, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20672158896923065, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20299996435642242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1855146437883377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10207734256982803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0976424515247345, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.1125464141368866, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.1042901948094368, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10323627293109894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09242185205221176, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08868436515331268, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05707255378365517, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.049678511917591095, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04868157207965851, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04844651743769646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028448278084397316, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02476746030151844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.024703441187739372, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022786008194088936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02275669015944004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014708384871482849, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014497282914817333, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014376888982951641, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009105179458856583, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028448278084397316, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.13.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21731485426425934, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19664745032787323, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18910449743270874, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16794651746749878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09887319803237915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09155214577913284, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11362985521554947, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10463504493236542, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.101082943379879, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.0857577994465828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08068561553955078, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05758732557296753, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04993491992354393, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04737860709428787, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04676719382405281, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02885049767792225, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024802550673484802, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024532847106456757, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022274751216173172, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021884283050894737, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.01555382739752531, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01583043485879898, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014761405065655708, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011145622469484806, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02885049767792225, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06969155371189117, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06501087546348572, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06315991282463074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05713879317045212, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.032312653958797455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03041953407227993, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.036610133945941925, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03389206528663635, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.032783083617687225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028840431943535805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0275226179510355, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018535302951931953, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01616513356566429, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015468360856175423, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015301624312996864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009262535721063614, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007973024621605873, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007904065772891045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0072740186005830765, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007207375019788742, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004835234954953194, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004892132245004177, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0046022506430745125, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003245176514610648, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.032783083617687225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06198358163237572, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05766702815890312, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05564387887716293, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05020203813910484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02840486727654934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.026552310213446617, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.032792892307043076, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.030414743348956108, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.028838591650128365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.025343388319015503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02422221750020981, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.016536494717001915, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014475922100245953, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013594081625342369, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013380872085690498, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008266721852123737, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0069828759878873825, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0068833790719509125, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0063592721708118916, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006255632732063532, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004314123652875423, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004297815263271332, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004041609354317188, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002769861137494445, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.032792892307043076, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1994926929473877, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18667282164096832, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18280760943889618, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16533724963665009, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09238437563180923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08777560293674469, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10337340086698532, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09484989941120148, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09361811727285385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08270566165447235, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07894373685121536, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.052324336022138596, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04516226798295975, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04404205083847046, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04377152770757675, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026064295321702957, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.022261880338191986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022179801017045975, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02024086005985737, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020185129716992378, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01334509626030922, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.012850278988480568, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012860293500125408, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007699980400502682, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026064295321702957, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19678600132465363, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17442536354064941, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.16669601202011108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13918781280517578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09023131430149078, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08158861100673676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10377924144268036, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09576650708913803, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09232772886753082, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07331183552742004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06561621278524399, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05301456153392792, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04596030339598656, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04346463084220886, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04285651072859764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026540342718362808, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02288813516497612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.022678017616271973, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019607003778219223, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019381331279873848, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014336161315441132, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014809566549956799, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013562094420194626, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01060437224805355, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026540342718362808, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16412873566150665, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1547602117061615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1519506424665451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.13891837000846863, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07659529149532318, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07323592901229858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08461806923151016, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07829969376325607, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07747279852628708, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06936612725257874, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06661394238471985, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04307843744754791, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.037448275834321976, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03668556734919548, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03650880604982376, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02152886614203453, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.019025791436433792, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.018976129591464996, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01759020797908306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.017563890665769577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011439592577517033, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011690158396959305, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011193234473466873, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008071479387581348, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02152886614203453, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2203165888786316, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20782503485679626, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20409563183784485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18647357821464539, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10271377116441727, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09823127090930939, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11326098442077637, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10495064407587051, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10387161374092102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09298994392156601, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08914215117692947, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05742292106151581, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.049986861646175385, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.048980213701725006, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04874534159898758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028638344258069992, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024906236678361893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.024842072278261185, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022910315543413162, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022880418226122856, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014810500666499138, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014547069557011127, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014475743286311626, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0090874545276165, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028638344258069992, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.14.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21881066262722015, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19823992252349854, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.19069750607013702, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16946294903755188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09963788837194443, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09234578162431717, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.1147284135222435, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.1054917722940445, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10191679000854492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08661836385726929, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08153308928012848, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.058245427906513214, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05048493295907974, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.047874242067337036, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.047249842435121536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029258714988827705, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02525581419467926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024979274719953537, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022771023213863373, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.022372348234057426, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015899719670414925, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.016380036249756813, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.015065365470945835, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011819728650152683, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029258714988827705, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07039012759923935, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06565014272928238, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0637887641787529, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.057688627392053604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03265441581606865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.030732017010450363, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03698963299393654, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.034244783222675323, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03312881290912628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.029141681268811226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027801858261227608, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018741419538855553, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01634548231959343, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01563715748488903, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015464983880519867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009370213374495506, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00806130189448595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007991285994648933, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007353048771619797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007286541163921356, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00489607872441411, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0049512023106217384, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004658290185034275, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003289920510724187, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03312881290912628, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05923468619585037, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05505507439374924, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05297389253973961, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04779317229986191, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.027151886373758316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.025273488834500313, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0315045565366745, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.029232734814286232, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0275651216506958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02419223077595234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.023115426301956177, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.015883060172200203, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.013934786431491375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012986164540052414, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012766201980412006, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007935728877782822, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006690084468573332, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006582432892173529, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0060966103337705135, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00598103879019618, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004160070791840553, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004160602577030659, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.003877955488860607, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002711082808673382, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0315045565366745, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2115166038274765, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19810420274734497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19402146339416504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1759779155254364, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.098445363342762, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0935254916548729, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10982728004455566, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10096391290426254, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09970913827419281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08821329474449158, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08428998291492462, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05561939626932144, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04809081554412842, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04694155231118202, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.046655528247356415, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027749327942728996, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023694206029176712, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02361294999718666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021572891622781754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021512504667043686, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014176777563989162, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013628479093313217, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01367868296802044, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00810218695551157, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027749327942728996, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17588192224502563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1447000503540039, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1337839961051941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11065894365310669, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07892999798059464, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06615005433559418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09437040984630585, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08667989075183868, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08272340893745422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05977429077029228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05503968521952629, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04839734360575676, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.041686829179525375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03817293047904968, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.037317417562007904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024283478036522865, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020323744043707848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02007528766989708, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01669991947710514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016262749210000038, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013267455622553825, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01364240050315857, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012161791324615479, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009930954314768314, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024283478036522865, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16592881083488464, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.15647411346435547, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15364177525043488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14043499529361725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07750388234853745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07411308586597443, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.0855800211429596, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.07921561598777771, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07840418815612793, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07018177956342697, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06741462647914886, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.043551769107580185, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.037891436368227005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.037120457738637924, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03694109246134758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02175416424870491, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01926160417497158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.019214121624827385, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.017809662967920303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.017782920971512794, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01153197418898344, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011853745207190514, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011281873099505901, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008204853162169456, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02175416424870491, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22150763869285583, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2089369148015976, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.205251544713974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1875268667936325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.1033458560705185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09883879125118256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.1139783263206482, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10559668391942978, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.1045321598649025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09356773644685745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08973123133182526, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05783800408244133, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05031880736351013, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04931284859776497, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04907156527042389, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028843427076935768, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02510688081383705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025039374828338623, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023101938888430595, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023070452734827995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014950528740882874, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014717038720846176, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014616153202950954, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00927208736538887, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028843427076935768, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.15.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21749615669250488, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19667890667915344, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18887117505073547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.1678466498851776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09890575706958771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09143882244825363, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11445598304271698, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10509474575519562, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10123361647129059, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08581767976284027, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08094394207000732, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05804060027003288, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05015737935900688, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.047441862523555756, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04678245261311531, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029162079095840454, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02487209439277649, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024581244215369225, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02234836108982563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021927829831838608, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015895074233412743, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015955189242959023, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.015032081864774227, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011270828545093536, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029162079095840454, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07443100214004517, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06949448585510254, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0676238089799881, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06126689910888672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03454308956861496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.032565925270318985, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03900003433227539, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03611232712864876, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03502822294831276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030883383005857468, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029488852247595787, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019774019718170166, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017245881259441376, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016533812507987022, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016365738585591316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009880677796900272, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008522883988916874, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008456616662442684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00778656592592597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007722577080130577, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005161770153790712, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00522496597841382, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004924121778458357, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003474961034953594, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030883383005857468, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06649475544691086, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.061921995133161545, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0598255917429924, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.0540606714785099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03051704913377762, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02854461967945099, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03512188792228699, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0325615257024765, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030942073091864586, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02724919095635414, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026073351502418518, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01771450787782669, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015490932390093803, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014585721306502819, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014364044182002544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008851195685565472, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007478080224245787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007378540467470884, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006817338988184929, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006711615715175867, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004619041923433542, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00458297785371542, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004336310550570488, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002940339967608452, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0325615257024765, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20290130376815796, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19001097977161407, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1860238015651703, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16861581802368164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09402088820934296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08937513083219528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10493496060371399, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09653284400701523, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09525265544652939, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0842493548989296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0804770216345787, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05315294861793518, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04596220701932907, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04479637369513512, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04453849047422409, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026502404361963272, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02267272211611271, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022590525448322296, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020637433975934982, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02059245854616165, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013566087000072002, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013102040626108646, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013091442175209522, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007877080701291561, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026502404361963272, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20066480338573456, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1760525107383728, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.16738778352737427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.14061462879180908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09107037633657455, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08131295442581177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.1067122146487236, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09825114905834198, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09412045776844025, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0745926946401596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06747124344110489, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05459034442901611, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04723482206463814, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04401379078626633, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04322223737835884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02738306298851967, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023421635851264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02316719852387905, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.02028290182352066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019951755180954933, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014902878552675247, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015576242469251156, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013907709158957005, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011424215510487556, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02738306298851967, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17198430001735687, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16223275661468506, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15926754474639893, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14552873373031616, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08040449768304825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07689516991376877, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.0888000950217247, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08216977119445801, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08132964372634888, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07279137521982193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06990727037191391, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04521021619439125, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03932284563779831, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.038529057055711746, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03834474831819534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02259034849703312, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020019153133034706, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.019968705251812935, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01851283758878708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018487893044948578, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012014445848762989, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012357600964605808, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011758383363485336, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008604517206549644, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02259034849703312, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22447502613067627, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21174629032611847, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20791754126548767, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1900414377450943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.1047416478395462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10018552839756012, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11556195467710495, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10701772570610046, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10593023896217346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09484068304300308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09095807373523712, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05864749848842621, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05099970102310181, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.049979958683252335, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04973562806844711, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02925924025475979, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025440596044063568, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025373568758368492, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023410223424434662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023378074169158936, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015191490761935711, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01490845251828432, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014851782470941544, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009387568570673466, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02925924025475979, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.16.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.22327621281147003, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.201711505651474, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.19363774359226227, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.17216961085796356, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.10149882733821869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09378010779619217, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11734533309936523, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10785138607025146, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.1039329320192337, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08802527189254761, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08300638943910599, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.059403423219919205, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05147750675678253, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04867665469646454, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04799783229827881, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02978573739528656, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.025495894253253937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.025202665477991104, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022895125672221184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.022462965920567513, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.016054198145866394, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.016319887712597847, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01515220943838358, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011498617008328438, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02978573739528656, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07320143282413483, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0684020146727562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06652959436178207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06033795326948166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.034011367708444595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03206038102507591, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03844781219959259, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.035607725381851196, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03449500724673271, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030436934903264046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02909189462661743, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019491761922836304, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017000563442707062, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016291089355945587, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016119007021188736, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009743712842464447, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008411826565861702, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008342515677213669, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007692928891628981, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00762595608830452, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005103351082652807, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005176999606192112, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004873921163380146, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003461756743490696, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030436934903264046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06462091207504272, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06024700775742531, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0581713542342186, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05260414257645607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.029698863625526428, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02779000997543335, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.034192051738500595, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03169746696949005, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03011261112987995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026528194546699524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025372983887791634, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01726064272224903, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015099850483238697, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014204598031938076, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013985863886773586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008626466616988182, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007298349402844906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007198757026344538, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006659358739852905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006560086738318205, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0045150150544941425, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004497738089412451, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004242850001901388, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00291584525257349, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03169746696949005, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21891938149929047, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2052178531885147, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20110855996608734, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1824694275856018, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10187196731567383, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09697264432907104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11373425275087357, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10448188334703445, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10316558182239532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09140362590551376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08746320009231567, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05761076509952545, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04973553866147995, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04856495559215546, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.0482778362929821, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0287175215780735, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02453603781759739, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02445824071764946, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02236286550760269, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022306062281131744, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01469266414642334, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014136750251054764, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01418780256062746, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008447068743407726, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0287175215780735, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1967410296201706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16872844099998474, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15860296785831451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.13152246177196503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08912280946969986, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07802072912454605, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10556624084711075, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09694655239582062, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09223159402608871, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07033074647188187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06388730555772781, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05405113473534584, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04664456099271774, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.043071795254945755, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.042201731353998184, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027161652222275734, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02296118251979351, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.022659949958324432, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.019398324191570282, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018993152305483818, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014901497401297092, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015366094186902046, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013802819885313511, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011258738115429878, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027161652222275734, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16957350075244904, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.15996181964874268, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15709960460662842, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1436082422733307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07924649119377136, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07580336183309555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08750218898057938, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08098825812339783, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08015251159667969, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07177998125553131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06895655393600464, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.0445268452167511, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.038748230785131454, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03796859085559845, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.037780921906232834, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022246103733778, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.0197030957788229, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.01965407468378544, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01822570338845253, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01819673739373684, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011798816733062267, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012132517993450165, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011544755659997463, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008409608155488968, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022246103733778, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2210434526205063, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2085370272397995, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20482376217842102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18722984194755554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10316124558448792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09870113432407379, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11373124271631241, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10539384931325912, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10433696955442429, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09344352036714554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08963014930486679, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.057677362114191055, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.050223708152770996, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04921932518482208, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04898591339588165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028766470029950142, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02504684217274189, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02497909776866436, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023050937801599503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023020708933472633, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014868396334350109, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014658358879387379, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01453591138124466, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009200765751302242, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028766470029950142, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.17.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21729223430156708, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19579622149467468, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18763510882854462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16692034900188446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09874071180820465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.0910012498497963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11447295546531677, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10526229441165924, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10114120692014694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08547576516866684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08060172945261002, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05796308442950249, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.050195951014757156, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.047335222363471985, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04664944112300873, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029093438759446144, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02477596327662468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.0244704931974411, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02221222221851349, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021770799532532692, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015773005783557892, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01585693471133709, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014877530746161938, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011116564273834229, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029093438759446144, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07240783423185349, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06778310239315033, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06603573262691498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.059953946620225906, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.033653225749731064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03182292729616165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.037952400743961334, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03510599583387375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03411013260483742, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0301923006772995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02890368178486824, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01922925002872944, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01675533317029476, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016103368252515793, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01594947837293148, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009611057117581367, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008290254510939121, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008229121565818787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0075937011279165745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007536525372415781, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005024068057537079, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005059871356934309, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004813880193978548, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003348573576658964, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0301923006772995, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06683510541915894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06240186467766762, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06039997190237045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.054725587368011475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030715400353074074, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.028860213235020638, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03511622175574303, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.032538916915655136, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03115837462246418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027522657066583633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026351198554039, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017727870494127274, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015487908385694027, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014667941257357597, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014471153728663921, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00885172188282013, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007491281721740961, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.0074038757011294365, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0068386453203856945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006752933841198683, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004599506966769695, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004526982083916664, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004351123236119747, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0028554245363920927, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.032538916915655136, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21633481979370117, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2033088207244873, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1992790549993515, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18116863071918488, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10074453800916672, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09597065299749374, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11219152808189392, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.103180430829525, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10200615227222443, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09060067683458328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08680963516235352, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.056838296353816986, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.049133166670799255, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04803380370140076, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.047774020582437515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02833990752696991, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024262117221951485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02417965605854988, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022143233567476273, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022123213857412338, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014501170255243778, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01395635586231947, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014028235338628292, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008323241025209427, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02833990752696991, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1852097064256668, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15763822197914124, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14604975283145905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11933846026659012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08413878083229065, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07203689962625504, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10167776793241501, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09393775463104248, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08690707385540009, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06540723890066147, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05936961621046066, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05203847959637642, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.045204032212495804, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04078200086951256, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03969091922044754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02610805071890354, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.021910691633820534, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02146991528570652, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01842408999800682, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017856959253549576, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014242175966501236, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015042534098029137, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012881597504019737, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011118147522211075, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02610805071890354, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1687091886997223, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.15922896564006805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1563636213541031, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14304117858409882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07890589535236359, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07549066841602325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08707774430513382, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08065059036016464, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07979883253574371, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07150880247354507, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06867596507072449, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.044325537979602814, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03860688954591751, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03783271461725235, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03764759749174118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022138457745313644, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01969531923532486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.01964094303548336, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018236519768834114, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01821018010377884, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011759699322283268, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01220956351608038, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011505335569381714, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008561993017792702, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022138457745313644, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22019270062446594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2078438103199005, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20417921245098114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18671879172325134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10277244448661804, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09834226220846176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11332151293754578, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10499071329832077, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10392303764820099, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09315142780542374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.0893615186214447, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05749589577317238, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05003276839852333, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04903087764978409, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04879602789878845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028666455298662186, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024957386776804924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.024892667308449745, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.0229828841984272, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022952090948820114, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01483415998518467, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014616133645176888, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014501048251986504, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00919317826628685, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028666455298662186, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.18.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21450795233249664, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1928151398897171, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18435056507587433, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16368155181407928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09746511280536652, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08945555984973907, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11327015608549118, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10437498241662979, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.0999186560511589, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08406709879636765, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07927677035331726, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05744149163365364, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.049840036779642105, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04675780609250069, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.045997247099876404, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02882903255522251, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024492302909493446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02417205460369587, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021905969828367233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02144598960876465, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015521924011409283, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01575208455324173, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014575335197150707, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011085761711001396, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02882903255522251, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0685863047838211, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06418067961931229, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.062465619295835495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05673365667462349, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03184570372104645, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.030073799192905426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.035977382212877274, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03328688070178032, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03227703645825386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028555095195770264, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027355119585990906, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018237760290503502, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015888262540102005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01524377055466175, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0150916064158082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009118341840803623, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007860707119107246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007799274288117886, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00720174377784133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00714368000626564, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004777045920491219, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004820211324840784, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004569645971059799, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0032091266475617886, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03328688070178032, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06864327192306519, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06405990570783615, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.062047116458415985, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05619138106703758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.031589679419994354, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.029656508937478065, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03611614182591438, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.033467959612607956, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03202218562364578, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0282805897295475, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027069613337516785, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018227633088827133, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01593887433409691, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015089635737240314, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01488224882632494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009111644700169563, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00772137101739645, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007628173101693392, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0070532504469156265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006961860693991184, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004750950261950493, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004693623166531324, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004491080529987812, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002991596469655633, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.033467959612607956, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20488324761390686, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1922149509191513, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18834365904331207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1711728423833847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09531789273023605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09067665785551071, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1062951311469078, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09787499159574509, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09650903940200806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08568741381168365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08202103525400162, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.053846005350351334, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04661652073264122, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04546216130256653, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04517035558819771, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026843693107366562, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02299405448138714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022900860756635666, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020978938788175583, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02091868221759796, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013736538589000702, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01328551210463047, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013269659131765366, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008000841364264488, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026843693107366562, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19483885169029236, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16581790149211884, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15617574751377106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1252092719078064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08842182159423828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07744251191616058, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10337210446596146, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0949874296784401, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0914209708571434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06772678345441818, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.060016341507434845, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0530923530459404, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04595748335123062, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.043028656393289566, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.042317308485507965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026757722720503807, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.023367932066321373, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02315138280391693, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01957228034734726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019293315708637238, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014957617036998272, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016060015186667442, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014070410281419754, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012398682534694672, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026757722720503807, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16907164454460144, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1596882939338684, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15684431791305542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14346005022525787, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07901716232299805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07561895251274109, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.0872236043214798, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0807570144534111, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07990527898073196, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07165716588497162, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06884397566318512, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04434177279472351, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.038631632924079895, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03784813731908798, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03766214847564697, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022140363231301308, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01964302733540535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.019591625779867172, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018182415515184402, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018155748024582863, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.0117079708725214, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012089803814888, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011455385945737362, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008381049148738384, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022140363231301308, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.21741245687007904, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20539161562919617, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20175838470458984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1846121996641159, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10141920298337936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09708687663078308, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11191123723983765, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10361342132091522, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10256537795066833, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09198779612779617, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08834785968065262, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.0567164421081543, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04938003048300743, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04838985949754715, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04815676435828209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028362974524497986, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02465532347559929, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02458902634680271, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02272244170308113, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022691864520311356, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014810528606176376, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014467309229075909, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014488465152680874, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009146001189947128, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028362974524497986, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.19.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21391937136650085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19158786535263062, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18236197531223297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.1620357483625412, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.0972045511007309, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08866370469331741, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.1152164414525032, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10523021966218948, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.0998239517211914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08383210003376007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07931426167488098, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.0586438849568367, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.050553832203149796, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.0469399020075798, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04604513943195343, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02993597649037838, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02510901354253292, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02470370940864086, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02260138839483261, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.022045303136110306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.0169810950756073, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01687038503587246, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01586868241429329, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.012506376951932907, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02993597649037838, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07584633678197861, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07106088101863861, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06914681196212769, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06285364180803299, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.035279758274555206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0333293080329895, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03988717868924141, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03691103681921959, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03574639558792114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.031683553010225296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03036942332983017, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02023053541779518, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017642280086874962, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016907742246985435, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01673305779695511, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010121379047632217, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008749931119382381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008678906597197056, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008031544275581837, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0079647870734334, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0053159575909376144, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005416312720626593, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005079552531242371, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0036663650535047054, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.031683553010225296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06528259068727493, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0610208660364151, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.0587935596704483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05328048765659332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030007723718881607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02804494835436344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03477592393755913, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03228093311190605, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03042505495250225, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026927294209599495, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02582293190062046, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01755017228424549, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01538983453065157, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014355813153088093, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014110892079770565, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008776512928307056, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007382756564766169, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007261511869728565, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00675850547850132, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0066331662237644196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004607880022376776, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004574785940349102, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004299477208405733, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002955879783257842, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03228093311190605, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2145981639623642, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2016066312789917, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1974390149116516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17981354892253876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09986644983291626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09510277956724167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11151368170976639, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10246780514717102, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10109812766313553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08989699184894562, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08616497367620468, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05642419308423996, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04879540577530861, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04760117083787918, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04732579365372658, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028147682547569275, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024034352973103523, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023942697793245316, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021955735981464386, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02188495732843876, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01437894906848669, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013826556503772736, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01387772522866726, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00821566954255104, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028147682547569275, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.19912408292293549, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16459305584430695, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.15313811600208282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1219690665602684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0890539139509201, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07587935030460358, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10547425597906113, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09677072614431381, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09314347058534622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06684589385986328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.059290990233421326, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0539129301905632, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04642535001039505, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04295298829674721, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.042100418359041214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02714158594608307, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.022736597806215286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02250652201473713, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018512286245822906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01812194474041462, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014965367503464222, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015023015439510345, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013906232081353664, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010806216858327389, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02714158594608307, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1733047366142273, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1636982560157776, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16080205142498016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14714473485946655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08090180903673172, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07744649052619934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08925756067037582, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0826396495103836, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08181192725896835, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07339715957641602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07053985446691513, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04538073390722275, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.039467502385377884, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03868035972118378, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03849676251411438, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022648965939879417, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01994020864367485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.019887147471308708, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.0184356477111578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01840849779546261, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.0119229881092906, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012073307298123837, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011665336787700653, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008130648173391819, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022648965939879417, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22088658809661865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2086840718984604, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20505166053771973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18763433396816254, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10301438719034195, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09863084554672241, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11352689564228058, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10520852357149124, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10416775196790695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09345706552267075, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08970515429973602, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05757860094308853, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05011393502354622, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.049132779240608215, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.048896417021751404, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028699064627289772, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024973420426249504, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02491028793156147, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02300957590341568, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022983139380812645, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014828006736934185, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014574495144188404, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014502196572721004, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009098581969738007, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028699064627289772, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.20.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21544858813285828, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19280627369880676, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18389824032783508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.163676455616951, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09767535328865051, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08930328488349915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11436052620410919, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.1050092950463295, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10033320635557175, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08416689187288284, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07955048978328705, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05793773755431175, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.0500977598130703, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04684840887784958, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04606541246175766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029102995991706848, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02454390749335289, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024195563048124313, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021932659670710564, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021429477259516716, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015777813270688057, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01580183394253254, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014748475514352322, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011064311489462852, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029102995991706848, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07217060774564743, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06763824075460434, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06575430929660797, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.059770528227090836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.033590007573366165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03170591965317726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0381133109331131, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03525080531835556, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03402964025735855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030174126848578453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.028945259749889374, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01934276893734932, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0168609656393528, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016105985268950462, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0159243606030941, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009678821079432964, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00835583545267582, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008280704729259014, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00767561374232173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007604035548865795, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005097999703139067, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005208970513194799, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004856552928686142, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0035548326559364796, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.033590007573366165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0640287697315216, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05990326404571533, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05768436938524246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05225657671689987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.02947166934609413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.027514979243278503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.034318823367357254, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03184687718749046, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.029876938089728355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0264875665307045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025430575013160706, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017286011949181557, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015169285237789154, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014097190462052822, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013843769207596779, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008640260435640812, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007238291669636965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007114205975085497, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006634687539190054, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00650717644020915, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00452509056776762, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004481207113713026, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004211528226733208, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002877580700442195, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03184687718749046, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2035929560661316, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19131815433502197, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18736028671264648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17059889435768127, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0946241244673729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09006092697381973, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10580772161483765, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09719589352607727, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09574844688177109, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08522327244281769, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08173811435699463, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05364237353205681, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04629598185420036, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.045095644891262054, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04480231553316116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02672726847231388, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02277045138180256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022680221125483513, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02082390896975994, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02076154388487339, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013673046603798866, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013128265738487244, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013147060759365559, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00783621333539486, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02672726847231388, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.21398748457431793, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.17599256336688995, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.16267164051532745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.12930595874786377, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.09568176418542862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.08091846108436584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11460120975971222, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10526145249605179, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.10012529790401459, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.07132011651992798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.06457866728305817, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05851495638489723, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0506468266248703, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04633444920182228, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04528050124645233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.029475033283233643, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.024892212823033333, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.024572569876909256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.020372727885842323, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.019851237535476685, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.016340242698788643, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016955837607383728, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.015053724870085716, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01260384637862444, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.029475033283233643, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17605337500572205, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16625666618347168, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16331970691680908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1494724452495575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08214900642633438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07862916588783264, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09071371704339981, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08391878753900528, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08306334912776947, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0745273232460022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07160758227109909, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04611365497112274, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04008788615465164, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.039293788373470306, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.039107296615839005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023034418001770973, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02029244787991047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020238567143678665, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01876457780599594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01874035969376564, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012169161811470985, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012333991937339306, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01191369816660881, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008360323496162891, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023034418001770973, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22447116672992706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21205559372901917, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20834767818450928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19061440229415894, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10473854094743729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10027537494897842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11547204852104187, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10693193972110748, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.1058863028883934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09501781314611435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09126576781272888, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.058576542884111404, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05096781998872757, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04996521770954132, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04973211884498596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029214046895503998, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025453802198171616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025391835719347, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02346392720937729, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02343703620135784, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015170694328844547, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014943094924092293, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01484228577464819, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009454869665205479, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029214046895503998, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.21.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2172972410917282, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19472503662109375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18585336208343506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16559191048145294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09852554649114609, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09020109474658966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11526288092136383, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10591118037700653, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10118089616298676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08509087562561035, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08045513927936554, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05837558954954147, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05055444687604904, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04725679010152817, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04645366966724396, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029460497200489044, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024745315313339233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024385884404182434, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02214382402598858, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021633222699165344, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.016160473227500916, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01591469906270504, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01515130139887333, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011120930314064026, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029460497200489044, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07260636240243912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06804843246936798, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06618130952119827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06018970161676407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03380054235458374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03191426396369934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03832655027508736, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0354432575404644, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.034249257296323776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03037913702428341, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029143987223505974, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01942373253405094, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01693093031644821, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016197357326745987, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016020745038986206, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00971700344234705, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008376977406442165, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.00830220989882946, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00769057497382164, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007618998177349567, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00510370684787631, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0051779900677502155, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004871891811490059, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0034848698414862156, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03037913702428341, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06549602746963501, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.061214786022901535, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05900737643241882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05352003127336502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030167004093527794, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02822517789900303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03499353304505348, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03240496292710304, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.030594464391469955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0270773246884346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025999054312705994, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.017625728622078896, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01542640384286642, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014435475692152977, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014192930422723293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008818629197776318, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007407419383525848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007293513510376215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006781691685318947, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006660583429038525, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004615341313183308, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004558703862130642, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004318052437156439, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0029264981858432293, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03240496292710304, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21200710535049438, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19914698600769043, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19509099423885345, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17741502821445465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09862368553876877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09388300776481628, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10996846854686737, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1012454554438591, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0998498722910881, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08877800405025482, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08504657447338104, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05570033937692642, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.048214614391326904, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.047019846737384796, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.046736814081668854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027769912034273148, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02374248392879963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023655224591493607, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02169794961810112, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021622924134135246, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014196503907442093, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013663817197084427, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013701238669455051, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008145046420395374, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027769912034273148, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17291384935379028, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.14397704601287842, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1351720690727234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1080363467335701, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07746139168739319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06680885702371597, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08994045108556747, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08287528902292252, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0809689313173294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05832955241203308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05251406505703926, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.045917607843875885, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.039718594402074814, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03727681562304497, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0366915762424469, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022989554330706596, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01960533857345581, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01948469504714012, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015924183651804924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015682891011238098, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012426020577549934, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012699936516582966, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011669360101222992, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009022539481520653, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022989554330706596, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17704270780086517, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16720162332057953, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16422057151794434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15037056803703308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08271869271993637, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07918127626180649, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09139204770326614, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0844699814915657, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08362316340208054, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07506178319454193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07217860221862793, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.046501047909259796, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04038233682513237, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.039587534964084625, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.039402466267347336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023247724398970604, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02048425003886223, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020431065931916237, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01895187795162201, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018931737169623375, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012356637977063656, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012511390261352062, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012098516337573528, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008558238856494427, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023247724398970604, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2233094424009323, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21098369359970093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20731423795223236, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18978555500507355, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.1042194813489914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09979335963726044, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.1148834377527237, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10643424838781357, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10538200289011002, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09457682818174362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09084905683994293, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.058283574879169464, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.0507163479924202, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04972716420888901, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04948824271559715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029063966125249863, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025302838534116745, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025238098576664925, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023323019966483116, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023292139172554016, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015047783963382244, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01480509340763092, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014724766835570335, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009297877550125122, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029063966125249863, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.22.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21401292085647583, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1913449913263321, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18211323022842407, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16244521737098694, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.0969315618276596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08834362775087357, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11385751515626907, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10485182702541351, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09966067224740982, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08365098387002945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07922252267599106, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.057554442435503006, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.049984753131866455, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04647446423768997, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04562105983495712, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02878209576010704, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024275850504636765, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.0238934513181448, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021694814786314964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021144410595297813, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015359505079686642, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015582859516143799, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014250967651605606, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010758991353213787, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02878209576010704, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0677632987499237, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06348748505115509, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0616498664021492, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.056036461144685745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.031559187918901443, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0297564510256052, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03596733883023262, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.033261366188526154, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03198450803756714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028353070840239525, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027225889265537262, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018272364512085915, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.015908636152744293, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015151336789131165, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014974008314311504, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009167936630547047, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007871275767683983, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007793615572154522, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007232369855046272, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007155909668654203, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0048658824525773525, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004928009118884802, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004627474583685398, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0033727854024618864, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.033261366188526154, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.061848562210798264, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.05783938616514206, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05563642829656601, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05044080689549446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.028503796085715294, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.026599686592817307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.033181559294462204, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.030789049342274666, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02889067493379116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.025579791516065598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02451998181641102, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.016727572306990623, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014674928970634937, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013642429374158382, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013392696157097816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008374983444809914, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007019871845841408, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00690071564167738, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006426374893635511, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006304611451923847, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.00439685583114624, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004363012034446001, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0040955133736133575, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002826359588652849, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.033181559294462204, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2081349641084671, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19546517729759216, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19133403897285461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17402739822864532, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09688808768987656, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09218808263540268, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10824764519929886, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09964951127767563, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09810563176870346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08711149543523788, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0835464745759964, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05485773831605911, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04746449738740921, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04619721695780754, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04588213562965393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027368267998099327, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023352429270744324, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02324710786342621, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021332405507564545, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021251913160085678, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013999833725392818, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013501359149813652, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013481438159942627, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008082415908575058, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027368267998099327, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16810862720012665, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13676168024539948, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12139074504375458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09871592372655869, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07450637966394424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.060153018683195114, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09758131951093674, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08890790492296219, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07866498827934265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.056860145181417465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0533020943403244, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04985402151942253, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.043138258159160614, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03673378750681877, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03507860004901886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025351468473672867, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02062300406396389, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019978778436779976, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01757231168448925, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01666903682053089, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014482058584690094, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015459857881069183, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012645188719034195, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.012130326591432095, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025351468473672867, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1777045875787735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16786156594753265, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16495157778263092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1510378122329712, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08301296830177307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07948160171508789, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09170963615179062, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08480094373226166, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08394953608512878, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07535327970981598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07246594876050949, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04665793851017952, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04051440209150314, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03970790654420853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.039521344006061554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023310331627726555, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020497389137744904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02044372446835041, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018956340849399567, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018933435901999474, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01233779639005661, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012443273328244686, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012081163004040718, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008421426638960838, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023310331627726555, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22158800065517426, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20931340754032135, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20575004816055298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1884012222290039, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10342638194561005, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09904272109270096, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11408042907714844, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10563206672668457, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10456983745098114, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0938870757818222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09021240472793579, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05790916457772255, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05034378916025162, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04935784265398979, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04912377893924713, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028889011591672897, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025134848430752754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025068091228604317, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023176956921815872, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02314595691859722, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015019569545984268, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014740237966179848, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014696178026497364, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009304802864789963, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028889011591672897, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.23.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21241183578968048, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.190419003367424, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18139246106147766, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16228316724300385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09627994894981384, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08796121925115585, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.1129375770688057, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10416541248559952, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09900476038455963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08347328752279282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07922806590795517, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.057148437947034836, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04967164993286133, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04615018889307976, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04528910666704178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028697215020656586, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02407122403383255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023673372343182564, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.0215766504406929, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021017838269472122, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015408135950565338, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015408235602080822, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014320118352770805, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010560148395597935, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028697215020656586, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06682000309228897, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06264347583055496, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06086556985974312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05535098910331726, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.031079357489943504, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029324375092983246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03530989587306976, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.032688066363334656, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03149542957544327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.027943922206759453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02681749500334263, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01788927987217903, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01560690626502037, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014882274903357029, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014707771129906178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.008949286304414272, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007677221670746803, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007602513302117586, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007045006845146418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0069742705672979355, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0046919179148972034, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004725496284663677, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004468298051506281, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0031391449738293886, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.032688066363334656, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0676111951470375, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06331416964530945, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.061227019876241684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.055584412068128586, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03121848590672016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.029288113117218018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03589183837175369, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0332995243370533, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.031642474234104156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02804478257894516, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026907028630375862, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01810622401535511, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015849612653255463, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014922319911420345, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014703597873449326, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00904618576169014, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00763833150267601, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00753206480294466, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006997402757406235, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006890745367854834, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0047306171618402, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004659573547542095, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004450847394764423, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0029617936816066504, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0332995243370533, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.19801145792007446, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1860874593257904, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18225710093975067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16598060727119446, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09212915599346161, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08771275728940964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10293985903263092, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09478089958429337, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09332476556301117, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08294159919023514, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0795488953590393, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.052100565284490585, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04510439187288284, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04390399530529976, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.043639469891786575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02596339024603367, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.022183328866958618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02209128998219967, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02028103917837143, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02020089328289032, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013280795887112617, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01280009001493454, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0127950981259346, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0076429229229688644, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02596339024603367, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.2001306116580963, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15799932181835175, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.14244444668293, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.10943663120269775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08951233327388763, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.07255914062261581, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10854048281908035, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09986384958028793, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.09394249320030212, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.06183389574289322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.056470710784196854, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0555841401219368, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04808967933058739, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04344131797552109, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04229317232966423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027957657352089882, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02338932640850544, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.02301778271794319, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018333304673433304, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.017721684649586678, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015393831767141819, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01610778644680977, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.013972374610602856, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011990123428404331, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027957657352089882, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17833107709884644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16851256787776947, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16556166112422943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15149778127670288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.083315908908844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0797608271241188, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09201072156429291, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08508671075105667, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08424317091703415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07560130208730698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07271444797515869, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.046803008764982224, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04065777733922005, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03985536843538284, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.0396624431014061, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023389071226119995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02056986093521118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020517827942967415, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019022732973098755, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018995974212884903, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012364201247692108, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01248199213296175, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012102724984288216, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008443417027592659, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023389071226119995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2219846099615097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20978128910064697, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20616845786571503, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1887541115283966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.1036243885755539, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0992271900177002, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11428485810756683, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10581561177968979, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.1047787293791771, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.094060517847538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09037479013204575, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05799764767289162, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05043397098779678, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04944619908928871, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.049209725111722946, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028931468725204468, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025171982124447823, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025107400491833687, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023207785561680794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023177722468972206, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015021286904811859, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01475144736468792, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014698285609483719, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00928441807627678, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028931468725204468, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.24.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21185098588466644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18950888514518738, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18006499111652374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16139119863510132, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09591862559318542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08730268478393555, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11302263289690018, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10431745648384094, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09871848672628403, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08309323340654373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07895360887050629, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.057162173092365265, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04971526563167572, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.045995358377695084, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.045090559870004654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028582610189914703, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02400924079120159, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023600371554493904, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02151244506239891, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.020937558263540268, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015212264843285084, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015427730046212673, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014042858965694904, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.01060747355222702, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028582610189914703, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07482918351888657, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07010582834482193, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06826236844062805, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.062074288725852966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03484172001481056, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03292756900191307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03939540684223175, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03644882142543793, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.035308416932821274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.031297992914915085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030031688511371613, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019983315840363503, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017435211688280106, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01671944558620453, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016548428684473038, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010004103183746338, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008688762784004211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008620286360383034, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007985247299075127, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007920539006590843, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005283389240503311, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00542591605335474, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0050583817064762115, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0037334770895540714, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.031297992914915085, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07055233418941498, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06605701148509979, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.063931405544281, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.057950299233198166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.032560452818870544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03054281324148178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03742591291666031, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.034657541662454605, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03301496058702469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02925349585711956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02804141864180565, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01890217512845993, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016540655866265297, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015575324185192585, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015339476987719536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009457291103899479, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007983447052538395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007875539362430573, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007306608371436596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007203604560345411, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004939403384923935, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0048804269172251225, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004649240989238024, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031268049497157335, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03301496058702469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20738400518894196, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19470439851284027, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19062498211860657, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17340274155139923, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09631185978651047, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09165724366903305, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10747651755809784, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09899886697530746, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09757757186889648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08662725239992142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08300895243883133, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05438724532723427, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04714201018214226, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04591059684753418, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04562368988990784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02713502198457718, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02318021096289158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02309134230017662, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021173253655433655, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021104099228978157, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013865405693650246, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.01335108932107687, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013371720910072327, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007937134243547916, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02713502198457718, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.20356199145317078, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.15022875368595123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.129399836063385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1033267006278038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08675423264503479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06442417204380035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.11141879111528397, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.10285119712352753, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0954110398888588, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05993185192346573, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05674291402101517, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05671992525458336, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04910271242260933, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04179390147328377, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03991039842367172, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02835756540298462, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02193227782845497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.021412918344140053, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01681184582412243, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.015555999241769314, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015027596615254879, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014920730143785477, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012692266143858433, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009951593354344368, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02835756540298462, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18355928361415863, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1734449416399002, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17038826644420624, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15597613155841827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0857481062412262, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08207515627145767, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09471456706523895, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08758030831813812, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.086702860891819, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07781314104795456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07481425255537033, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.048134587705135345, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04182063415646553, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04099312797188759, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.040796324610710144, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02405114844441414, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021101584658026695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.021046273410320282, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01950579509139061, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01947936788201332, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012682474218308926, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012723542749881744, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012413113377988338, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008500191383063793, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02405114844441414, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22551734745502472, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21312938630580902, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20941711962223053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19174453616142273, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10530076920986176, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10082638263702393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.1160753071308136, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10753754526376724, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10646789520978928, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.095584936439991, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09179177135229111, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05889497697353363, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05124232918024063, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05023197829723358, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.049988023936748505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029366247355937958, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025545191019773483, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02548014000058174, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023552294820547104, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023522624745965004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015197967179119587, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014935415238142014, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014864303171634674, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009351436980068684, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029366247355937958, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.25.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21629691123962402, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19337724149227142, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18384873867034912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16496005654335022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09782370924949646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08909934759140015, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.115323506295681, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10627426952123642, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.1007426455616951, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08488781005144119, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08071213215589523, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05838547274470329, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05060448497533798, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04686088114976883, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.045947037637233734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02926289290189743, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02437746152281761, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023962847888469696, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021843116730451584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021256329491734505, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.01569681614637375, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015546604990959167, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014512578025460243, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010537163354456425, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02926289290189743, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07307946681976318, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06858056783676147, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0667625442147255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.060785066336393356, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.034049347043037415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03221186622977257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0385979525744915, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03562740981578827, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03449798747897148, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030660802498459816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.029475897550582886, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.019602065905928612, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017051976174116135, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016348334029316902, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016180720180273056, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.00983081478625536, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008509278297424316, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008441146463155746, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007837115786969662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00777089037001133, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005237284116446972, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005333803128451109, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0050190044566988945, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0036944374442100525, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030660802498459816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06901222467422485, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06468864530324936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06249481812119484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05678170546889305, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03186144307255745, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.029930220916867256, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03668789565563202, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03396806865930557, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03231311962008476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02865106426179409, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027531087398529053, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018537547439336777, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016186580061912537, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015257395803928375, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015032985247671604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009268886409699917, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007849650457501411, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007741788402199745, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007205587346106768, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007100019138306379, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004874357953667641, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004844876937568188, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004595108330249786, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00316044595092535, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03231311962008476, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20222577452659607, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18995080888271332, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1861198991537094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1693873554468155, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09389679878950119, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08940505981445312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10463477671146393, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09653399139642715, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09508093446493149, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0846242904663086, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08103184401988983, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.052928730845451355, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04593023657798767, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04474787041544914, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04446978121995926, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026388704776763916, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02259385585784912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022494923323392868, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020651698112487793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020587142556905746, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013481991365551949, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013006838038563728, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013024481013417244, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007723588030785322, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026388704776763916, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1999615579843521, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.16148026287555695, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1480911374092102, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.11191750317811966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.08898881077766418, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.074098140001297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10740526765584946, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.09817871451377869, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0932704508304596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0642387717962265, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05679124966263771, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.05491413176059723, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0473143570125103, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.04321404919028282, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.04221406579017639, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027875661849975586, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0233572069555521, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.023045159876346588, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.018830513581633568, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.018349360674619675, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.015832828357815742, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.016082147136330605, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.014637695625424385, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.01211733091622591, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.027875661849975586, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18713679909706116, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17682082951068878, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17375193536281586, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15908323228359222, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08738510310649872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08367728441953659, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09643812477588654, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08925206959247589, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08835870772600174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07931786775588989, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07623805105686188, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04896219074726105, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.042603567242622375, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04176424816250801, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04157048091292381, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024464379996061325, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02146705985069275, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.0214088074862957, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01983707584440708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019810615107417107, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012847909703850746, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012898207642138004, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01257265079766512, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008557374589145184, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024464379996061325, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22655229270458221, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2141222059726715, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.21043458580970764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19266445934772491, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10578722506761551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10131184756755829, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11665000766515732, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.1080232709646225, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10695929825305939, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0960334986448288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09222723543643951, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05917046219110489, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.051489029079675674, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05047567933797836, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.05023813620209694, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029518919065594673, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025686603039503098, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025623410940170288, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023683825507760048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02365516871213913, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015302585437893867, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.015042854472994804, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014971409924328327, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009464769624173641, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029518919065594673, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.26.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21669504046440125, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1935339868068695, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18370282649993896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16486918926239014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.0979737862944603, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08914428949356079, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.1164039671421051, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10683280974626541, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10094291716814041, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08498811721801758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08083602786064148, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05874256044626236, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05091698467731476, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04702696204185486, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04607275500893593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029515616595745087, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024613339453935623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024187827482819557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022095415741205215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.0214960016310215, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015974588692188263, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01591663248836994, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014764131046831608, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011041006073355675, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029515616595745087, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07508974522352219, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07046432793140411, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06858029961585999, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06247049942612648, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03494756668806076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03305024281144142, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.039521437138319016, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.036596134305000305, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03540584444999695, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0314507782459259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030170777812600136, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020016811788082123, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017458850517868996, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01672990992665291, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016553090885281563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010009825229644775, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008618996478617191, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008544029667973518, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007915183901786804, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007845387794077396, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005241960287094116, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005278044380247593, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005009518004953861, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0034949886612594128, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0314507782459259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06860854476690292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06429456174373627, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.062097154557704926, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05644278600811958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.031706392765045166, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.029731467366218567, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.036515019834041595, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.033908404409885406, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.032120365649461746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02851821482181549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027359576895833015, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018423745408654213, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0161475520581007, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015162650495767593, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014915446750819683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009217576123774052, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007756311446428299, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007642267271876335, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00710265664383769, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0069907489232718945, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004800359718501568, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0047310274094343185, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004507198464125395, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002991083776578307, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.032120365649461746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21236300468444824, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.199563130736351, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19571208953857422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17847517132759094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09918470680713654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09442193806171417, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11062613129615784, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10190968215465546, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10041549056768417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0893266499042511, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08570947498083115, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05601640045642853, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04853039234876633, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.047292228788137436, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04699593782424927, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02794390171766281, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023895885795354843, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023799030110239983, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021847601979970932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021787676960229874, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014287467114627361, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013770434074103832, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013778623193502426, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008219730108976364, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02794390171766281, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1622881144285202, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13570313155651093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12656232714653015, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1041378602385521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07318625599145889, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06249383091926575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08642100542783737, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07940831035375595, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07585682719945908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05620192363858223, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05062688887119293, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04407312348484993, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.038050685077905655, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03523017838597298, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.034541890025138855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022151190787553787, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018542688339948654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01831519789993763, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015263373032212257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014924384653568268, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012081916444003582, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012094463221728802, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011209020391106606, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008540838956832886, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022151190787553787, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17758016288280487, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1678168773651123, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16492733359336853, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15103137493133545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08292292803525925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07939787954092026, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09151356667280197, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08469545096158981, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08383233845233917, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07526355981826782, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07244081050157547, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04651914909482002, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.040433451533317566, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03963607922196388, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.039448902010917664, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023225590586662292, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02040005475282669, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020345360040664673, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01885593868792057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.0188289787620306, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012221797369420528, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012291050516068935, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011962314136326313, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008209886029362679, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023225590586662292, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.21615365147590637, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20431728661060333, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20076729357242584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1839047223329544, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10088694840669632, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09662692248821259, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11121749877929688, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10303377360105515, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10201158374547958, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09159690141677856, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08805287629365921, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05643041804432869, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.049107521772384644, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04813787713646889, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04791023209691048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028135478496551514, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024499008432030678, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.024437077343463898, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022590994834899902, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022556878626346588, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014571244828402996, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014340751804411411, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014255156740546227, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009017976000905037, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028135478496551514, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.27.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.20151065289974213, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.17942191660404205, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.1698712855577469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15220370888710022, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09100548923015594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08240867406129837, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10836435109376907, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.09976684302091599, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09391088038682938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07868881523609161, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07482711970806122, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05480381101369858, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.0475713275372982, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04369046911597252, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.042739249765872955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027469754219055176, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.022879768162965775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.022449661046266556, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.020491721108555794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.019882524386048317, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.01473724003881216, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.014852290973067284, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.013504823669791222, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010273795574903488, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027469754219055176, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0685231015086174, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06427790224552155, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.062480729073286057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05691460147500038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03192253038287163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.030136093497276306, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03626652806997299, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.033576011657714844, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03235173225402832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028735123574733734, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027606064453721046, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018377132713794708, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0160381980240345, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015302719548344612, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.015125807374715805, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009194605052471161, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007907547056674957, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007831573486328125, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007266385480761528, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007194309029728174, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004830969963222742, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004888155497610569, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0046050348319113255, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0032781108748167753, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.033576011657714844, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.061970315873622894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.058086097240448, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05596442520618439, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05092458054423332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.028654951602220535, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.026798782870173454, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0333721823990345, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.030902419239282608, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.029043007642030716, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.025776222348213196, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.024842457845807076, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01684361882507801, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014714207500219345, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013717779889702797, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013473371043801308, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008437959477305412, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007039067335426807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006922677159309387, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006462655961513519, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006338693201541901, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004422198981046677, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.0043501597829163074, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004126318730413914, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.002792699495330453, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0333721823990345, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21286700665950775, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20014317333698273, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19620060920715332, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17898836731910706, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.099318727850914, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09462806582450867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.1107083410024643, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10187741369009018, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1005919948220253, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08948216587305069, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08586189895868301, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05608134716749191, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04854978621006012, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04736855626106262, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04710008203983307, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02797100320458412, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023904265835881233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023824511095881462, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021849146112799644, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02178771048784256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01428331807255745, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013727754354476929, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013792773708701134, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008131680078804493, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02797100320458412, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17624808847904205, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11905880272388458, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.09056106954813004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.083749920129776, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07401067763566971, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04456775635480881, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.10323521494865417, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.095244862139225, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0830252468585968, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05040787160396576, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05218179151415825, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.052781421691179276, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.045759428292512894, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03623585030436516, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03362799808382988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.026423310860991478, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019819241017103195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01897885464131832, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015618236735463142, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01384025625884533, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014348363503813744, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014870874583721161, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01138779055327177, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010750005953013897, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03362799808382988, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17638881504535675, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16673991084098816, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16386161744594574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15012870728969574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08233773708343506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0788482055068016, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09089131653308868, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08410269767045975, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08325157314538956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07477155327796936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07196157425642014, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04616468399763107, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04013853892683983, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03934745863080025, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03916409611701965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023043448105454445, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020229501649737358, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02018076926469803, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018705731257796288, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018677998334169388, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012096350081264973, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012166060507297516, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011840138584375381, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00808550976216793, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023043448105454445, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2142070233821869, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2025163322687149, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.19902975857257843, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18235106766223907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0999613106250763, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09576612710952759, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11018924415111542, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10208162665367126, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10106353461742401, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09082154929637909, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08733251690864563, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05594131350517273, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04865332692861557, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.047706957906484604, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04748403653502464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02789703942835331, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024302512407302856, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02423929050564766, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022422943264245987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022398142144083977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014483047649264336, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01426782924681902, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014168898575007915, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009024417027831078, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02789703942835331, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.28.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.20116424560546875, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.17977114021778107, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17008762061595917, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15256062150001526, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09106837958097458, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08254707604646683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10911504924297333, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.1002110168337822, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.0938967764377594, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07906994968652725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07528136670589447, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05520037189126015, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.0479796938598156, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.043920740485191345, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.042920954525470734, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02784343808889389, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023344682529568672, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02288767881691456, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021069185808300972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02043900266289711, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015219464898109436, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015608101151883602, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01391882635653019, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011289360001683235, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02784343808889389, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0672423243522644, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06302851438522339, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.061210453510284424, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05572027713060379, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03126534819602966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.029480591416358948, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0355481393635273, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03296336904168129, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.031675003468990326, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028115572407841682, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02700669690966606, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01801670528948307, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01572638750076294, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014979146420955658, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014797508716583252, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009021648205816746, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.0077284243889153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007648773491382599, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0070927925407886505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007016586139798164, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.004736423958092928, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004764098674058914, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004505496472120285, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0031673379708081484, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03296336904168129, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06347081810235977, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.059392064809799194, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.057295508682727814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.052044034004211426, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.029298899695277214, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.027423759922385216, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03401266410946846, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03152411803603172, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.02968430146574974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.026333188638091087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.025309497490525246, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01716340333223343, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015015210956335068, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014010325074195862, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013770759105682373, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008579111658036709, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007177654653787613, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007058839313685894, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006574708968400955, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006454857531934977, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004482661373913288, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004414605442434549, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004182614851742983, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0028060779441148043, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03152411803603172, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20153455436229706, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18913385272026062, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1853838413953781, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1687808781862259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09376433491706848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08920355886220932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10469546169042587, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09643474221229553, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09499330073595047, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08445751667022705, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08096001297235489, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05302858352661133, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04592224955558777, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.044717978686094284, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04441632702946663, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02642112597823143, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02257593162357807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02248082496225834, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020639877766370773, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.020566005259752274, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01349908858537674, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013002283871173859, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013022836297750473, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007720788940787315, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02642112597823143, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16948255896568298, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13366082310676575, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1184346005320549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0977439358830452, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07519429922103882, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05883744731545448, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09624070674180984, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08772224187850952, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07956176996231079, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05574600398540497, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.051736511290073395, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04919951409101486, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04247939586639404, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03688660264015198, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03546171635389328, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025013936683535576, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02048143930733204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019970549270510674, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.017080169171094894, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.016271991655230522, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.014262236654758453, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.015019425190985203, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012618442066013813, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011694593355059624, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.025013936683535576, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17585451900959015, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1662784218788147, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16338394582271576, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14974868297576904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0820758119225502, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07861270010471344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09063171595335007, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0838114470243454, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08300112932920456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07457004487514496, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07179508358240128, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04605851322412491, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.0400116965174675, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03922789543867111, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03904157504439354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023000773042440414, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020152144134044647, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020102161914110184, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018639428541064262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018612833693623543, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012087749317288399, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012097056023776531, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011829373426735401, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0080160116776824, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023000773042440414, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.21318891644477844, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20158833265304565, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1981453001499176, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18167400360107422, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0994940921664238, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09532681107521057, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10976705700159073, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10160069167613983, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10059716552495956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09041320532560349, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08704417943954468, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05572890490293503, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04843376949429512, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04748914763331413, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.047263942658901215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02780231088399887, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02419397607445717, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.024130983278155327, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022329654544591904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02230175957083702, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014460165053606033, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014201601967215538, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014148798771202564, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008989743888378143, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02780231088399887, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.29.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2007591277360916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.17956262826919556, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.16971156001091003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15246734023094177, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09085281193256378, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08228129893541336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10945896804332733, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10037822276353836, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09371194243431091, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07912420481443405, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.075440414249897, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05538735166192055, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04805837944149971, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04381045326590538, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04276331514120102, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.0280921533703804, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02327149361371994, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02279689721763134, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02104157954454422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.020389318466186523, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015626247972249985, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015568576753139496, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014355858787894249, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011225124821066856, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.0280921533703804, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07745428383350372, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07249128818511963, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07049811631441116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0640852227807045, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03607933223247528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.034019146114587784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04090746492147446, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03784311190247536, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.036575786769390106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.032378796488046646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03105025365948677, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02078690566122532, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018152449280023575, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01733892410993576, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017144758254289627, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010435473173856735, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00905962847173214, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008980670012533665, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008331582881510258, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008252632804214954, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005563496612012386, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0057366834953427315, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005313308909535408, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004024077672511339, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.032378796488046646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06684207916259766, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06251423060894012, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06027922034263611, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.054719842970371246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03083559311926365, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02881007082760334, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03577645123004913, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03310765326023102, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0312659852206707, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02765553444623947, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.026567043736577034, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01804046891629696, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01578449085354805, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014748875983059406, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014502769336104393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00902728270739317, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007575869094580412, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007455163169652224, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00693876389414072, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006812192965298891, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004734492395073175, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004677149001508951, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004425044637173414, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0030082426965236664, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03310765326023102, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.19899551570415497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.18627788126468658, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18230946362018585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.16566264629364014, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09218215197324753, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.08756484091281891, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10310310870409012, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09500228613615036, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09340599924325943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08275870233774185, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.07929811626672745, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05215708538889885, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04517798125743866, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04392724856734276, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.043612901121377945, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0260081235319376, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.022167466580867767, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.022078167647123337, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.020205531269311905, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02013733983039856, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013282549567520618, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.012803441844880581, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.012793047353625298, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007631325162947178, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0260081235319376, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.15159957110881805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11635731160640717, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1048373281955719, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08559123426675797, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0676748976111412, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05283091962337494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08026904612779617, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07331319898366928, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07049478590488434, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.046355511993169785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.042991891503334045, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04107419028878212, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03603611886501312, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03357992693781853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03298431634902954, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.021409831941127777, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01927570439875126, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019115779548883438, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015798570588231087, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0155277568846941, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012983323074877262, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014333921484649181, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012337351217865944, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011990205384790897, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03357992693781853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17983326315879822, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1700155884027481, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16704711318016052, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15306314826011658, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0839790552854538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08041158318519592, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09270608425140381, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08577992767095566, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08490680903196335, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07626277208328247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07336215674877167, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04708404839038849, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04094332084059715, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.040127817541360855, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.0399361290037632, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023523546755313873, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02058805711567402, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020536581054329872, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019024554640054703, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018997222185134888, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012336128391325474, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012316983193159103, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01207047887146473, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008102853782474995, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023523546755313873, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.21920613944530487, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20724789798259735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.2036512941122055, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18669281899929047, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10232774168252945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09797780215740204, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11285272985696793, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10450899600982666, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10344061255455017, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0929194986820221, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.0893540307879448, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05725663900375366, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04982416331768036, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04882548004388809, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.048589129000902176, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028583426028490067, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024865727871656418, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.024801278486847878, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022940531373023987, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022908568382263184, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014855004847049713, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014588306657969952, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014530078507959843, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009210326708853245, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028583426028490067, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.30.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.20514468848705292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18530775606632233, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17642748355865479, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15864422917366028, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.0931331068277359, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.0853605642914772, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11033911257982254, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10158315300941467, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09575984627008438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08172926306724548, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07777777314186096, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05585501343011856, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04856235906481743, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04481103643774986, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04388606175780296, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028221402317285538, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023608606308698654, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023186910897493362, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021414151415228844, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02083379589021206, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015532153658568859, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015472710132598877, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014406785368919373, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.01094610895961523, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028221402317285538, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07809804379940033, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07339870184659958, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07152066379785538, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06522749364376068, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.036415476351976395, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03448178246617317, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04113660380244255, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03804338723421097, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03687544912099838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.032824255526065826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.0315483957529068, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02084307000041008, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018144026398658752, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017422854900360107, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017251119017601013, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010421788319945335, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008946719579398632, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008875107392668724, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008216654881834984, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.00815148837864399, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005445962306112051, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005427780095487833, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005219717044383287, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0035404847003519535, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.032824255526065826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06787600368261337, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06371285766363144, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06168048456311226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.056127727031707764, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03141196817159653, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02951209247112274, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.036170151084661484, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03340234234929085, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03180249035358429, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.028292424976825714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027208635583519936, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01826208271086216, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015903456136584282, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014999373815953732, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014782468788325787, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009121964685618877, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0076537905260920525, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007548030000180006, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.00701447669416666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006913369987159967, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0047434307634830475, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004628657829016447, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0044671716168522835, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0028929978143423796, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03340234234929085, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20440880954265594, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19214844703674316, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1883496791124344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17181982100009918, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09515156596899033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0906389132142067, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10583120584487915, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09763043373823166, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09630434215068817, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08576861768960953, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08224627375602722, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.053564250469207764, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.046485789120197296, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.045324958860874176, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04506046324968338, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026689011603593826, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02287302352488041, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02278898097574711, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02091878093779087, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02087051421403885, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.013638691045343876, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013132017105817795, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013200923800468445, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007777731399983168, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.026689011603593826, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1631360650062561, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.1278003752231598, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.1099548488855362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09344582259654999, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07067946344614029, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05386146157979965, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0960211232304573, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08825194835662842, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07674261182546616, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05444609001278877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05168268084526062, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.049217816442251205, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.042472608387470245, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03463716059923172, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.032544661313295364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02477683499455452, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019024532288312912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018252165988087654, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.016100099310278893, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014823410660028458, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013809417374432087, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.014149351976811886, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011467237025499344, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010473478585481644, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.032544661313295364, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1829887330532074, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17295405268669128, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1699703335762024, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15569105744361877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0854673758149147, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08181487768888474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.0943564921617508, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08730535954236984, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08642987906932831, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07758637517690659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.0746111273765564, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.047932155430316925, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04167189076542854, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04082973301410675, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.040632009506225586, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023938309401273727, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020947547629475594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020889025181531906, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01934751495718956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01932203210890293, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.0125586474314332, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012519643642008305, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012287658639252186, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008226059377193451, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023938309401273727, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22273318469524384, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21060602366924286, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20691633224487305, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18959666788578033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10401777178049088, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09958136081695557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11472142487764359, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10625911504030228, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10518482327461243, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09445102512836456, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09081585705280304, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05822450667619705, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05065854266285896, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.049646928906440735, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04940399155020714, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02903996780514717, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025292376056313515, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025227034464478493, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02333107590675354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02329733781516552, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01506967656314373, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014854112640023232, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014739484526216984, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009397045709192753, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02903996780514717, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.31.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21158240735530853, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19178925454616547, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18284514546394348, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.1646156907081604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.0961458832025528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.088360995054245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11379146575927734, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10471711307764053, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09866505861282349, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08465617150068283, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08056763559579849, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.057609956711530685, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.049928754568099976, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04611724242568016, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.045180074870586395, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029039951041340828, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024057267233729362, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.023625394329428673, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02181832119822502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021228022873401642, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015851682052016258, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01545725017786026, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014687527902424335, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010586226359009743, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029039951041340828, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.080553337931633, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0756322368979454, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0736733078956604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06709751486778259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.037608757615089417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03557632863521576, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04242326319217682, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03927728533744812, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.038102347403764725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03383363410830498, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03245668485760689, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02150062844157219, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018755923956632614, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01800941489636898, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017834819853305817, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010752039030194283, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009277427569031715, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009204820729792118, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008517752401530743, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008450670167803764, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0056244186125695705, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005672117229551077, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005393320228904486, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003761202795431018, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03245668485760689, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06801058351993561, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06375545263290405, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06159067153930664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05599399656057358, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03148392215371132, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.029527224600315094, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03657074272632599, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03376421332359314, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03189273178577423, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02833539992570877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0272914320230484, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018476853147149086, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01607438363134861, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015069437213242054, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014828956685960293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009236742742359638, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007734081707894802, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007618841715157032, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007096509449183941, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006978278048336506, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004830098710954189, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004753728862851858, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0045203096233308315, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0030467675533145666, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03189273178577423, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22738692164421082, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2134883850812912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20921632647514343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.19052152335643768, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10602450370788574, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.10086912661790848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11815810948610306, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.108857162296772, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10741954296827316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09541212767362595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.09144650399684906, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0598415844142437, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.051861003041267395, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.05057626590132713, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.05023952201008797, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02983703836798668, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02553054690361023, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.025437964126467705, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.023313164710998535, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.023247018456459045, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.015242347493767738, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014683039858937263, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014730442315340042, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008729629218578339, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02983703836798668, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09512897580862045, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0766955241560936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07121556252241135, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05585598945617676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04268162325024605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03565247729420662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.049591172486543655, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.045807480812072754, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.044457875192165375, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030863367021083832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027786793187260628, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02523614838719368, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021866656839847565, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020471647381782532, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.020130109041929245, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012619567103683949, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01060933992266655, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010521398857235909, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008291616104543209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008135710842907429, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.00670983549207449, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006666271947324276, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006266117561608553, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004463220480829477, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.030863367021083832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18518882989883423, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1750555783510208, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17197784781455994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15756428241729736, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08650904893875122, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08281123638153076, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09559306502342224, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08839479833841324, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08747529238462448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07851389795541763, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07554974406957626, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04856352508068085, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.042174842208623886, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04132538288831711, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04112819954752922, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024277273565530777, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.0211506150662899, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02109357714653015, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019522275775671005, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01949252374470234, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012755641713738441, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012571175582706928, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012480358593165874, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0081614525988698, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024277273565530777, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2255118191242218, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21323221921920776, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.2095581442117691, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.191951185464859, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10538319498300552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10090259462594986, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11625909805297852, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10764671862125397, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.1065518707036972, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09566590934991837, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09198570996522903, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05904867872595787, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05133224278688431, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05029897391796112, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.050053246319293976, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.0294607225805521, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02562718838453293, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02555818110704422, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.023629428818821907, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02359492890536785, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015313811600208282, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01504622120410204, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014975141733884811, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009511147625744343, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.0294607225805521, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.32.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21370165050029755, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1944989114999771, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18557977676391602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16717015206813812, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09722086787223816, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08959554880857468, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11498337984085083, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10591623187065125, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09964350610971451, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08589115738868713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08165688812732697, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05789715424180031, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.050476837903261185, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04665197432041168, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.045710351318120956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028977273032069206, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024370549246668816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02392759546637535, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022165566682815552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021570108830928802, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015392398461699486, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015695203095674515, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01418489683419466, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010788947343826294, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028977273032069206, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07815801352262497, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07325708121061325, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07125567644834518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06483227014541626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.036445725709199905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.034417472779750824, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.041279859840869904, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03821704909205437, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.036945685744285583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03274606168270111, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.031408119946718216, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020930929109454155, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018234970048069954, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017450766637921333, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017261944711208344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010468544438481331, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008978324942290783, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008899763226509094, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008228477090597153, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008152619935572147, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005479585379362106, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005485175643116236, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005234749522060156, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0036088654305785894, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03274606168270111, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06892756372690201, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06449420005083084, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.062297165393829346, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05651753768324852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03185334801673889, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.029817374423146248, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.0369308777153492, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03414905071258545, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0322556346654892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0285892803221941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027499524876475334, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018634404987096786, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016259387135505676, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015231109224259853, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01497683022171259, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009322302415966988, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007796337828040123, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007676276843994856, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007133514620363712, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007020062766969204, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004871627781540155, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004767202772200108, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00455872667953372, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0030175072606652975, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0322556346654892, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21253660321235657, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1992526352405548, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19513414800167084, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17744185030460358, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09883850812911987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09394201636314392, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11024711281061172, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10173508524894714, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10012330114841461, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08876924216747284, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08494707942008972, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05580848082900047, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04842611774802208, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.047096606343984604, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04681318625807762, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027806002646684647, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023783544078469276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023687103763222694, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02169547602534294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021611716598272324, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014216354116797447, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013714849017560482, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013708161190152168, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008150818757712841, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027806002646684647, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16330376267433167, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12308067083358765, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10942748934030533, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08892673999071121, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07232465595006943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05503077805042267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08774159103631973, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08016910403966904, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07675452530384064, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.049859803169965744, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.045689672231674194, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04467903450131416, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03843791037797928, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03487890213727951, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03401140123605728, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022505197674036026, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01840553805232048, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018184347078204155, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014146200381219387, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013621375896036625, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012325381860136986, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012197556905448437, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011250106617808342, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008583122864365578, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022505197674036026, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18798954784870148, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17765402793884277, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17450252175331116, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1598326563835144, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0878666490316391, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08411172032356262, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09710638225078583, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0898122489452362, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0888672024011612, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0797213688492775, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07662198692560196, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.049333687871694565, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04285616800189018, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04198797792196274, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04178339242935181, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024642664939165115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021515507251024246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.0214548259973526, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01985219679772854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019823485985398293, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012928958982229233, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012821359559893608, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012647755444049835, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008360265754163265, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024642664939165115, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22869840264320374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2161611169576645, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.2124062478542328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19453753530979156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10692769289016724, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10235559195280075, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11798315495252609, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10925941914319992, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10812294483184814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09701193869113922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.0931987315416336, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.059941358864307404, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05211194232106209, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05105505511164665, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.05080438032746315, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029894404113292694, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02601572498679161, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025945156812667847, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02398012951016426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023943061009049416, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015530415810644627, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.015286301262676716, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015188122168183327, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009679690934717655, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029894404113292694, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.33.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2189328372478485, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19966696202754974, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.19091488420963287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.1720578372478485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09964589029550552, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09211290627717972, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11766282469034195, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10807308554649353, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10207603126764297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08816733211278915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08394056558609009, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05933550372719765, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.051433924585580826, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04771703481674194, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04680934175848961, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029792986810207367, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024749785661697388, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024322109296917915, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022490492090582848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02191237546503544, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015985747799277306, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01567254588007927, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014810379594564438, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010508677922189236, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029792986810207367, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07722543179988861, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07232388108968735, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07026392221450806, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06391967833042145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.036003101617097855, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03394360840320587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.040879737585783005, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03785437345504761, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03650912642478943, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03231288492679596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.030979052186012268, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020717328414320946, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018078848719596863, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.017254982143640518, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017057549208402634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010360670275986195, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.00889633595943451, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008814522996544838, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00815253984183073, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008070740848779678, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005433537065982819, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005472227465361357, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005176827311515808, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0036337929777801037, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03231288492679596, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06669500470161438, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06235506758093834, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.05996159091591835, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.054368358105421066, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.030782897025346756, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02868790179491043, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03617300093173981, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03341690078377724, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.031213566660881042, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.027630535885691643, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02658005617558956, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018228020519018173, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.015909045934677124, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.014732500538229942, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014445709995925426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009113797917962074, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007567276246845722, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007430337835103273, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006932525429874659, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006786682177335024, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004772091284394264, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004701853264123201, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0044226315803825855, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0030172676779329777, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03341690078377724, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21961772441864014, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20581269264221191, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.2015308439731598, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18320897221565247, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10219288617372513, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09708011895418167, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11416950076818466, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10508225113153458, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.1035776138305664, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09175772964954376, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0878060907125473, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05778944492340088, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.050058763474226, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.048747651278972626, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04842187091708183, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02880854159593582, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024623502045869827, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02452213130891323, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02244753949344158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022372936829924583, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014738615602254868, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014215826988220215, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014206419698894024, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008490466512739658, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02880854159593582, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16062979400157928, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12881813943386078, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11178139597177505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0884934514760971, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.06964855641126633, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05396139621734619, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09563286602497101, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08728282153606415, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07539437711238861, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.053377676755189896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04974638670682907, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.048631880432367325, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.041768573224544525, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03377310559153557, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03161945194005966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024510055780410767, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01802978105843067, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.017188454046845436, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014968621544539928, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013631324283778667, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013459427282214165, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01296056155115366, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011005941778421402, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008883829228579998, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03161945194005966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18778295814990997, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17743779718875885, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.174290731549263, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15962672233581543, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08789296448230743, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08410897105932236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09723339974880219, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08982600271701813, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08888301998376846, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07970134168863297, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07664993405342102, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04947669804096222, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.042895957827568054, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04202140495181084, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04181951284408569, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024737045168876648, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021572571247816086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.021518021821975708, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019910484552383423, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019881123676896095, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.013077197596430779, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012921375222504139, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012798738665878773, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00852112751454115, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024737045168876648, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2311568558216095, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21841885149478912, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.21460860967636108, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1965615451335907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10813239961862564, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10348179191350937, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11941777914762497, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.11049982160329819, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10934317111968994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09808171540498734, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09425724297761917, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.060703158378601074, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.052710652351379395, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05164237692952156, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.05139303207397461, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030287228524684906, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.0263226181268692, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.026249505579471588, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.024251027032732964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02421577274799347, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015800802037119865, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.015477683395147324, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015453129075467587, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00981617160141468, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030287228524684906, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.34.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2194075584411621, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.2006276398897171, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.19193795323371887, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.17318211495876312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09999015927314758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09257122874259949, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11741140484809875, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10844843834638596, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10236790776252747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.088721863925457, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08442658185958862, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05941104516386986, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.051587458699941635, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.047848720103502274, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.046936310827732086, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029754560440778732, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02475034072995186, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02432476542890072, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022527126595377922, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02195274829864502, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015858404338359833, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015588345006108284, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014693297445774078, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010349508374929428, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029754560440778732, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0692615956068039, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06479239463806152, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06267870217561722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.056881822645664215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03226026892662048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0302735548466444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.037139661610126495, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03431687504053116, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03271693363785744, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.028899183496832848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.027738073840737343, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.018820002675056458, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01641913130879402, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.015515413135290146, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01529785804450512, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009482916444540024, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008109715767204762, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008011427707970142, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007444801740348339, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007345857098698616, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005105337593704462, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005164677742868662, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0048316847532987595, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0035966485738754272, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03271693363785744, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.060497965663671494, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0564691461622715, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.054110486060380936, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04895975440740585, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.027853665873408318, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.025868484750390053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.032958999276161194, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.030447743833065033, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.028263531625270844, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02494724653661251, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02397737465798855, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.016620703041553497, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.014481076039373875, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.013337344862520695, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.013051502406597137, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.008311397396028042, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.006850630976259708, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006714428775012493, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006262345239520073, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.00611923960968852, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004359899088740349, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004275491461157799, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004012261051684618, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00273765716701746, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.032958999276161194, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20611043274402618, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19289623200893402, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.18870504200458527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17113801836967468, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0958741158246994, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09092095494270325, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10720199346542358, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.09876064956188202, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09716097265481949, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08583980798721313, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08205022662878036, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05430058389902115, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0470212958753109, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04567772150039673, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04538077488541603, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02708575874567032, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023110855370759964, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02300894446671009, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021039137616753578, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02094653621315956, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01385712530463934, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013382936827838421, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013333170674741268, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00801803171634674, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02708575874567032, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.164946511387825, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12194652855396271, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10457372665405273, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08502447605133057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07138481736183167, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.054505228996276855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09209822863340378, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08411996066570282, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0773240327835083, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04903244599699974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.048057977110147476, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04718828946352005, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04080429673194885, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.0351223461329937, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03366268798708916, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024157298728823662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019536618143320084, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01909666135907173, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015681201592087746, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014803851954638958, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013959604315459728, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01439547911286354, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012337022460997105, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.011153570376336575, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024157298728823662, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18850882351398468, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17802569270133972, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17484910786151886, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.16014927625656128, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08824896812438965, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08442241698503494, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09768263250589371, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09020707756280899, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08926725387573242, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08000698685646057, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.0769716128706932, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04966088384389877, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04308360069990158, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.042202457785606384, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04199371114373207, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024859609082341194, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021679958328604698, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02161920629441738, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.020002424716949463, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019972892478108406, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.013182159513235092, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.013002201914787292, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01290065050125122, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008593806996941566, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024859609082341194, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.23332077264785767, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.22039370238780975, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.21654793620109558, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19824697077274323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10918496549129486, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10446391999721527, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.12051567435264587, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.11158055067062378, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.11042347550392151, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0989716425538063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09508217871189117, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.061216093599796295, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.0532238595187664, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.0521390326321125, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.05187848582863808, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030551852658391, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.026571383699774742, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.026497457176446915, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.024474646896123886, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.024440491572022438, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015911035239696503, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.0156173100695014, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015557260252535343, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009892783127725124, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030551852658391, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.35.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.22183707356452942, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.20299573242664337, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.1941586583852768, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.17528760433197021, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.10114803910255432, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.0936407595872879, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.1189764067530632, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10988124459981918, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10354774445295334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08987035602331161, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08558284491300583, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.06014562398195267, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.052327536046504974, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.048419054597616196, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04746713861823082, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.030177252367138863, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.025070589035749435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024620503187179565, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022843917831778526, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.022249242290854454, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.016155648976564407, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015835780650377274, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014960471540689468, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010537008754909039, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.030177252367138863, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08126287907361984, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07587335258722305, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07365373522043228, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0668116956949234, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.037887271493673325, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03563091531395912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04307059198617935, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.039846692234277725, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.038441698998212814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03384435921907425, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03240039199590683, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021857745945453644, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.019056925550103188, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.018184971064329147, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.017977721989154816, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010938159190118313, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009443724527955055, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009359033778309822, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008639048784971237, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008554649539291859, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005785855930298567, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005905796308070421, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00551501614972949, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004042831715196371, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03240039199590683, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06783512234687805, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06329821050167084, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06080286204814911, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05504429340362549, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03131524845957756, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02913825958967209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03680911660194397, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03400999680161476, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.031767476350069046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.028038863092660904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02689841389656067, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018577855080366135, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016203097999095917, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01500558853149414, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014715111814439297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00930026639252901, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007727065589278936, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007588695269078016, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0070612807758152485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006918743252754211, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004890851676464081, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004820751957595348, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004526084754616022, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031304822769016027, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.031767476350069046, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21694490313529968, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20276539027690887, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19826306402683258, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1796865463256836, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10060186684131622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09532973170280457, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11250197142362595, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10367383807897568, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10201387107372284, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08992747962474823, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08597985655069351, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.056911394000053406, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04934224858880043, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.047950949519872665, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04761439189314842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028370993211865425, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024234075099229813, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024126669391989708, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022017788141965866, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021930746734142303, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014511361718177795, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014023647643625736, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013961349613964558, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008397206664085388, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028370993211865425, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14035481214523315, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11441847681999207, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10646931827068329, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08071933686733246, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.062292542308568954, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.052607592195272446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07286839187145233, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06709320843219757, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06533364206552505, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04416055232286453, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.039879463613033295, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03704417869448662, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03202773258090019, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.029848216101527214, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0293267834931612, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.018519623205065727, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.015533425845205784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.015420067124068737, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012010681442916393, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.011752349324524403, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009843113832175732, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00986644346266985, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009143989533185959, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00671441899612546, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03202773258090019, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18704089522361755, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1766209602355957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17347821593284607, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1588076800107956, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08765655755996704, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08383260667324066, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09694644063711166, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08962531387805939, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08865299820899963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07943076640367508, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07643654942512512, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04931769147515297, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.042830247431993484, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.041944414377212524, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.041735630482435226, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02468862012028694, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021577958017587662, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.021522480994462967, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019913004711270332, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019882172346115112, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.013070700690150261, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012996374629437923, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012790551409125328, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00865844078361988, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02468862012028694, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.23388084769248962, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.22091205418109894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.2169715315103531, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19854608178138733, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10950502753257751, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10475905239582062, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.12084804475307465, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.11192143708467484, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.11077632009983063, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09923453629016876, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09532744437456131, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.061385441571474075, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05339760705828667, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05230206251144409, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.052035778760910034, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030614444985985756, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.026638997718691826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.026568343862891197, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.024523142725229263, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.024489862844347954, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015875596553087234, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.015634354203939438, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015519275330007076, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009868953377008438, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030614444985985756, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.36.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2211565226316452, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.20243212580680847, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.19344238936901093, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.1747346818447113, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.10070529580116272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09322509169578552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11884728819131851, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10974923521280289, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10318206995725632, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08958472311496735, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08540017902851105, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.059944167733192444, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05213363096117973, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04815596342086792, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04717840254306793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029924917966127396, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024829454720020294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.0243655014783144, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022620709612965584, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.022000158205628395, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015684962272644043, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015565467067062855, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014398235827684402, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010138765908777714, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.029924917966127396, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08259150385856628, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0773354098200798, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07517517358064651, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06831061840057373, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03853090479969978, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.036324866116046906, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04380454495549202, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.040509894490242004, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03906549885869026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03455644100904465, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.033149056136608124, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022204605862498283, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.019342511892318726, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.018461890518665314, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018246980383992195, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011110116727650166, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009506762959063053, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009418516419827938, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008704147301614285, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008617579005658627, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005828355439007282, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005828461144119501, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005556114949285984, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003842635080218315, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.033149056136608124, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.0712384507060051, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06668861210346222, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06414276361465454, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05823560804128647, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.0329391248524189, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03073524311184883, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.038585562258958817, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0356636606156826, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03340104594826698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.029556311666965485, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.028424827381968498, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019416440278291702, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016964426264166832, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015762802213430405, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01546616479754448, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009707932360470295, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008064950816333294, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007923991419374943, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007369223516434431, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007225700654089451, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005059701856225729, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004955670330673456, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004699538461863995, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031276738736778498, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03340104594826698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22318138182163239, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20904937386512756, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20468614995479584, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18596890568733215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10381957143545151, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.0986131802201271, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11579472571611404, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10669459402561188, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10521689057350159, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0930822342634201, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08903910219669342, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05862459912896156, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.0508233979344368, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.049516722559928894, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.049191512167453766, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.029229754582047462, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.025026408955454826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024941403418779373, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022801220417022705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022731583565473557, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014977946877479553, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014482393860816956, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014452038332819939, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008709434419870377, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.029229754582047462, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.17916220426559448, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13828825950622559, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12495169043540955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09862682968378067, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07967673242092133, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06211913377046585, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09585244208574295, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08728234469890594, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08420221507549286, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05531693249940872, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.050863947719335556, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04893862083554268, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04195935279130936, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03851405531167984, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.037661612033843994, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024726994335651398, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.02051626518368721, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.0203261598944664, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015979401767253876, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0155452536419034, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013807134702801704, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013765646144747734, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012760642915964127, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010040546767413616, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024726994335651398, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1840478777885437, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1737191528081894, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17064857482910156, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15613336861133575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08626025915145874, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08248896896839142, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09531227499246597, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08817750215530396, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08724672347307205, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07812168449163437, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07507016509771347, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.048479609191417694, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04213328659534454, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.041258104145526886, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04105880856513977, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024229194968938828, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021206803619861603, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.021151186898350716, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01956157200038433, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019531700760126114, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012748295441269875, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012745829299092293, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01246922928839922, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008453533053398132, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024229194968938828, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2334621101617813, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.22041165828704834, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.21651937067508698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19814887642860413, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10934337973594666, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.1045743003487587, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.12080545723438263, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.11177192628383636, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.11058937013149261, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0990469753742218, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09511447697877884, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.06135350838303566, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.053328026086091995, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.052231188863515854, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.05197147652506828, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030631741508841515, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02662879228591919, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.026553025469183922, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.024510126560926437, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.024473682045936584, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.0159751009196043, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01566307432949543, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015610306523740292, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00993407517671585, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030631741508841515, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.37.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21077623963356018, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19290168583393097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.1838945746421814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16616685688495636, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09608063101768494, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08871018141508102, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11418982595205307, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.1052001491189003, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09843534231185913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08544921875, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08147300779819489, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.057485081255435944, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05014125630259514, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04609911888837814, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04509924352169037, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02882976457476616, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024035107344388962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.0235761646181345, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021956516429781914, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021351519972085953, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015416916459798813, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015465362928807735, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014190766960382462, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010569119825959206, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02882976457476616, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08512970060110092, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07980255037546158, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07768773287534714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07068471610546112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03977297991514206, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03758000582456589, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04490905627608299, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.041589170694351196, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.040311746299266815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.035695068538188934, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.034203264862298965, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022759515792131424, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.019841495901346207, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019034234806895256, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01884147711098194, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011368832550942898, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009778273291885853, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009698374196887016, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008947600610554218, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.0088747413828969, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005934907589107752, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005933136213570833, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005679760128259659, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0038660529535263777, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022759515792131424, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07241564989089966, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.067795529961586, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06540126353502274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.059312473982572556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.033495254814624786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03136656805872917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03897041827440262, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03594581037759781, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03396090492606163, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03005470335483551, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02885591611266136, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.01966739073395729, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01712530106306076, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016029972583055496, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015769433230161667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009833884425461292, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008212912827730179, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008084421046078205, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007507807575166225, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007380184717476368, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005140832159668207, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005033174529671669, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004802854731678963, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003200782695785165, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.033495254814624786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21204936504364014, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1987362951040268, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19467025995254517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1768074482679367, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09871935099363327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09378588944673538, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10996478796005249, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10148471593856812, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09998901188373566, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08856579661369324, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08469019830226898, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.0556408129632473, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.048314377665519714, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04706268757581711, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04674734175205231, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02773837372660637, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02378053590655327, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.02368338033556938, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021666765213012695, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02159777283668518, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01418394222855568, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013705343939363956, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013701855204999447, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008167916908860207, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.02773837372660637, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.18159306049346924, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13369455933570862, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11681260168552399, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.09381043165922165, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07977889478206635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.061272792518138885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0975649282336235, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08963248133659363, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.08504932373762131, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05248773843050003, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05011733993887901, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.049558963626623154, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04283886402845383, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03837484493851662, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03724685683846474, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024864859879016876, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.020076995715498924, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.019782911986112595, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014864357188344002, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01413416676223278, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013405817560851574, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01319978293031454, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.012020002119243145, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.008989174850285053, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.024864859879016876, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18488407135009766, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17452843487262726, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.17140409350395203, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1567470282316208, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08669416606426239, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08288144320249557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09580940753221512, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08862771838903427, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08768443018198013, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07848421484231949, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07533077150583267, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04870560020208359, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04234892129898071, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04147396236658096, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.041264962404966354, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02434438280761242, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021323001012206078, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02126573584973812, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01966201886534691, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01962946355342865, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012807393446564674, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012824579142034054, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012525036931037903, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008513393811881542, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02434438280761242, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.23317240178585052, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.22015798091888428, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.21620957553386688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1977626085281372, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10925470292568207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10445918142795563, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.1206100657582283, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.11167687177658081, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.11051078140735626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09891799092292786, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09491134434938431, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.06129922717809677, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05327964946627617, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05218249931931496, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.05192597582936287, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030582282692193985, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.026591524481773376, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02651789039373398, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.024464581161737442, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.024425875395536423, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015884965658187866, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.015622659586369991, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015528904274106026, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009884198196232319, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030582282692193985, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.38.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2241201251745224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.2053324431180954, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.1961439996957779, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.17728398740291595, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.10207748413085938, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09452396631240845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.12044531852006912, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.11138733476400375, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10458778589963913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.09090670943260193, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08671082556247711, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.060862403362989426, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.052901800721883774, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.048788830637931824, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.0477776825428009, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.03035942278802395, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02510012686252594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.024621594697237015, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02287478931248188, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.022234292700886726, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015863696113228798, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015673834830522537, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01452542468905449, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010104423388838768, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.03035942278802395, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09043542295694351, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08470074087381363, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08243333548307419, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07489284873008728, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04225940629839897, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0398828461766243, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04780103638768196, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04426281154155731, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04283057153224945, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.037900280207395554, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03627243638038635, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024253135547041893, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.021157244220376015, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020264288410544395, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02004973031580448, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012130809016525745, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010479366406798363, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010392103344202042, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009603967890143394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009517946280539036, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006380441598594189, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006473882589489222, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.006100941449403763, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004357974976301193, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024253135547041893, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07612255960702896, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07119500637054443, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06878924369812012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06244488060474396, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.035311732441186905, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03313582018017769, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.04084482043981552, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03773600235581398, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03582713007926941, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03165866807103157, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.030376458540558815, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020630808547139168, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01798515021800995, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.01692010834813118, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016650483012199402, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010328257456421852, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008685012347996235, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008564002811908722, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007939571514725685, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007815430872142315, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005396171938627958, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00533099053427577, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.005064758006483316, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003427132498472929, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03165866807103157, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2171158343553543, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20352111756801605, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1992625892162323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1810126006603241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10119523853063583, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09607948362827301, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11297894269227982, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10414639115333557, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10253333300352097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09074388444423676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08679485321044922, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05719706043601036, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04956856742501259, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04825014993548393, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04791665077209473, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0285041444003582, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0243915356695652, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024291977286338806, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.022218914702534676, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02215578779578209, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014594439417123795, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014099548570811749, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014083280228078365, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00846034288406372, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.0285041444003582, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13015925884246826, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09812816977500916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08814138174057007, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07094272971153259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05723406746983528, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.0450388602912426, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06857626885175705, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06273675709962845, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.061018653213977814, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03936532884836197, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03638458624482155, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.03494435176253319, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030088676139712334, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.027645956724882126, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02705366350710392, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01764068379998207, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.014744369313120842, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014635427854955196, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01142063271254301, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.011081964708864689, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009795715101063251, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00989489071071148, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00906350277364254, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007233879994601011, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.030088676139712334, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18263673782348633, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17232297360897064, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.169221431016922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1547660082578659, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08564789593219757, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08188324421644211, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09469953924417496, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08757708966732025, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08664164692163467, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0775289461016655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07444894313812256, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04816821217536926, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04185636341571808, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04098842293024063, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04078254848718643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024079490453004837, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02108364924788475, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.021029585972428322, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01944183185696602, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019411617890000343, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012696359306573868, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012701114639639854, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012415812350809574, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008463971316814423, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024079490453004837, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.23171307146549225, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21872226893901825, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.21480824053287506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19650240242481232, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10858754068613052, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.1038319319486618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.12042970955371857, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.11099288612604141, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10983943939208984, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09829222410917282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09458117932081223, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.06123785302042961, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05303318053483963, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.0519343726336956, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.0516829714179039, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030678726732730865, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.026613423600792885, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02654196321964264, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.024518420919775963, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02447909116744995, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.016259213909506798, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.015859944745898247, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015905026346445084, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.010340778157114983, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.030678726732730865, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.39.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21882256865501404, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.2000889927148819, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.19077478349208832, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.17244069278240204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09959769994020462, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.09200438857078552, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11866486072540283, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10927432775497437, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.10221866518259048, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08868663012981415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08466649055480957, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.0599251389503479, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05198343098163605, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04770839214324951, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.046662069857120514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.030004343017935753, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.02472700923681259, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.0242353193461895, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022574951872229576, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.02191147394478321, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.01595360040664673, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015729961916804314, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014591721817851067, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010448470711708069, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.030004343017935753, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08697035163640976, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0815424993634224, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07937610894441605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07214384526014328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04062700271606445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.038371872156858444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04587555676698685, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04252822697162628, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04115337133407593, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03647099435329437, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.034897420555353165, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023267343640327454, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.020306391641497612, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019455663859844208, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019251372665166855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011638816446065903, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010012738406658173, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009929083287715912, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00917208194732666, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009092810563743114, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006096185650676489, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.0061168186366558075, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005830937996506691, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004031452350318432, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023267343640327454, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07388056814670563, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06917442381381989, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06677955389022827, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.060623329132795334, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03426189720630646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03207411617040634, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03971431404352188, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03666553646326065, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03473041579127312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.030754998326301575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.029514076188206673, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020095180720090866, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.01749255135655403, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.0163832176476717, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01612056978046894, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.01003301702439785, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008398767560720444, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008271344006061554, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007682356983423233, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007557688280940056, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0052549890242516994, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005152106750756502, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004919037222862244, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032814480364322662, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.030754998326301575, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2194574922323227, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20590554177761078, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20176227390766144, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.18340572714805603, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10233090072870255, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09735412150621414, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11392825841903687, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10504360496997833, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10368704050779343, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.0918487012386322, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08786655962467194, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.057652220129966736, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05002044886350632, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04877391830086708, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.048509709537029266, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028753304854035378, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024645060300827026, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024549104273319244, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02245095930993557, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.022396346554160118, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014695419929921627, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014169531874358654, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014216732233762741, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00841727014631033, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028753304854035378, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1665855199098587, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12666425108909607, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.11171673238277435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0937475636601448, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07342808693647385, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05602334439754486, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.09186896681785583, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08354520797729492, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07799860090017319, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.052544429898262024, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04919371008872986, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04670056700706482, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.04010917991399765, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03554149717092514, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.0343879871070385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02363269403576851, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.019043458625674248, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018682103604078293, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015182996168732643, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01451998483389616, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.013244041241705418, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.013123446144163609, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011872655712068081, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009560149163007736, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02363269403576851, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1814451813697815, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.171156644821167, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16807295382022858, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1537177562713623, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08512909710407257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08136725425720215, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09406081587076187, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08702138066291809, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08610614389181137, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07705553621053696, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07394254207611084, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.0478547029197216, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.0415947288274765, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.040735285729169846, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04053376987576485, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.0239222664386034, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02095019817352295, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020893815904855728, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019315436482429504, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019286159425973892, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012602880597114563, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012611924670636654, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012330387718975544, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008387002162635326, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.0239222664386034, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22675515711307526, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.21400786936283112, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.21018019318580627, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.19227980077266693, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10631226003170013, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.10164138674736023, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11742126196622849, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.1086750403046608, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.1075543761253357, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09624281525611877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.09238776564598083, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05973881855607033, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.05186522752046585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.05080563575029373, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.05055500939488411, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02981960028409958, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02593223936855793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025857847183942795, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02386404760181904, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02382652834057808, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015579018741846085, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.015300081111490726, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.015234824270009995, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00976331066340208, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02981960028409958, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.40.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21090549230575562, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.19243961572647095, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18316349387168884, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16554175317287445, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09602035582065582, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08843692392110825, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11431967467069626, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10572999715805054, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09863881766796112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08534963428974152, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08152206242084503, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05791780352592468, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.05036943778395653, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.046065088361501694, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.045012637972831726, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02896747551858425, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023994779214262962, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02350117452442646, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.021901531144976616, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021236516535282135, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015326601453125477, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015441028401255608, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.013943730853497982, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.01044662855565548, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02896747551858425, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.09364798665046692, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08780335634946823, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08556321263313293, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07780417799949646, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04377356171607971, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.041413355618715286, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.049229469150304794, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.045649342238903046, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.04434525594115257, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.039283256977796555, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03759869560599327, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024982325732707977, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0218055322766304, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.020975589752197266, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02077535167336464, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.012489241547882557, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01082705706357956, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010745207779109478, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00992490816861391, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009852624498307705, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006556525826454163, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006647124420851469, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0062954924069345, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004451433662325144, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.024982325732707977, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07525141537189484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.07050704956054688, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06822071969509125, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06189423054456711, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.034870512783527374, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.032741621136665344, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.040250878781080246, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03713785856962204, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.035336386412382126, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.031287726014852524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.030041489750146866, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.020330797880887985, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017685305327177048, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016672316938638687, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016429925337433815, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.010157673619687557, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008518058806657791, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008404947817325592, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.0077861156314611435, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007675218395888805, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.0052874451503157616, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005168607924133539, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00497602391988039, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003259890479966998, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.031287726014852524, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.22003187239170074, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20658348500728607, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.20251302421092987, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.1842273771762848, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10266375541687012, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09770265221595764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11412067711353302, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10527098923921585, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10395239293575287, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.09226418286561966, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08830404281616211, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05779614672064781, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.05013510584831238, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04894644021987915, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04864998534321785, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028827277943491936, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024705449119210243, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024627402424812317, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02253354713320732, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.02248755656182766, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014727680943906307, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.014194082468748093, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.014278074726462364, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008424350991845131, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028827277943491936, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.1678563505411148, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.12235094606876373, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10782648622989655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.08764756470918655, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07303191721439362, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05438384413719177, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08811020106077194, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.08086443692445755, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0781722143292427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04874097555875778, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.04544522985816002, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.04481573402881622, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0387943759560585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03531467169523239, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03446117043495178, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022603025659918785, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.018975213170051575, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018806347623467445, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.01455245353281498, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.014025900512933731, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01250970084220171, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012953122146427631, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011454244144260883, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009587505832314491, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022603025659918785, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17849542200565338, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16843196749687195, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.1654127836227417, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.15122507512569427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.0837702676653862, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.08009929955005646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09255877137184143, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08563672006130219, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08474552631378174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07583030313253403, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07280555367469788, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04714271053671837, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04095865786075592, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.040119487792253494, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03991956263780594, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023567333817481995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020686576142907143, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02063334360718727, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01908598281443119, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019054358825087547, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012451369315385818, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012532097287476063, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012175577692687511, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008441454730927944, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.023567333817481995, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.22237814962863922, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20989584922790527, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.20616808533668518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18855434656143188, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10425477474927902, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09966594725847244, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11503923684358597, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10654076188802719, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10543607920408249, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.09436193853616714, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.090524822473526, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05846274644136429, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.050829797983169556, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04979637265205383, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.049555160105228424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029166143387556076, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.025386415421962738, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.025319218635559082, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.02335481531918049, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.023324904963374138, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.015157090499997139, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014930298551917076, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01481938362121582, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.00946441013365984, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.029166143387556076, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.41.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.21000193059444427, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.1908429116010666, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.18111222982406616, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.16365034878253937, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09542842209339142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08750160038471222, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11460797488689423, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10564085096120834, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09820254892110825, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08463640511035919, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.08082076907157898, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05801015719771385, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.050443921238183975, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04590526968240738, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04477648064494133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02930627390742302, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.024126626551151276, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02361360378563404, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.022043075412511826, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.021345824003219604, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.016030369326472282, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01585542783141136, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014645259827375412, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011059675365686417, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02930627390742302, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08825370669364929, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08268604427576065, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08057072758674622, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07317150384187698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.04120192676782608, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03895985335111618, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.046380601823329926, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04300558567047119, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.041763950139284134, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0369543619453907, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03533343970775604, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023510759696364403, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02051776833832264, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.019725311547517776, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.019534267485141754, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011755197308957577, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010138994082808495, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010060741566121578, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009275475516915321, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009202397428452969, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.006148474290966988, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.006158228497952223, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005897438619285822, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.004034135956317186, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.023510759696364403, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07385233044624329, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06907827407121658, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06695374846458435, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06068328768014908, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03424312546849251, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.03216755390167236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.039402298629283905, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03633708506822586, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0347239151597023, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03071320243179798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02944030798971653, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019907880574464798, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017327724024653435, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016374623402953148, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01614568755030632, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.00996123906224966, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008373523131012917, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008269059471786022, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007651717867702246, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007546248380094767, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005179828964173794, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005072879139333963, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004879982676357031, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003206258174031973, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03071320243179798, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21438272297382355, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.2010873556137085, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.1969483494758606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17896820604801178, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.10005491971969604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09517965465784073, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11139115691184998, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10274028778076172, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10134926438331604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08977154642343521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08579250425100327, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05634251609444618, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04893239960074425, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04773224890232086, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.047443125396966934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028085924685001373, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024097738787531853, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.024002600461244583, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021952304989099503, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021892521530389786, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01435948722064495, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013850554823875427, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013913610950112343, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008217382244765759, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028085924685001373, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.12213727086782455, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.08692096918821335, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07445669174194336, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06535571068525314, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.054262109100818634, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03807969391345978, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06618274748325348, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.060479965060949326, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05747605860233307, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03624206781387329, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.034717898815870285, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.033751267939805984, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02908533252775669, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.026324167847633362, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.02564360946416855, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017117198556661606, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01426312793046236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014079732820391655, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.011019202880561352, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01062877755612135, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.009689133614301682, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.009882859885692596, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.008882859721779823, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0074520716443657875, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.02908533252775669, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1760718822479248, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16611804068088531, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16316090524196625, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14919114112854004, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08261018991470337, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0789855495095253, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.0911775603890419, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08442721515893936, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08357444405555725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07477635890245438, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07174809277057648, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.046352799981832504, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.04035048559308052, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.039528582245111465, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.0393345020711422, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02316264435648918, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020322807133197784, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.020269962027668953, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01873588189482689, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.018710164353251457, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012153255753219128, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01221873052418232, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011890378780663013, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008113891817629337, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02316264435648918, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.21786059439182281, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.20565631985664368, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.201990008354187, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18470337986946106, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10211065411567688, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09764489531517029, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11268977075815201, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10432523488998413, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10326577723026276, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0924205482006073, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08869042247533798, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05727221816778183, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.049791619181632996, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04877754673361778, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04853527620434761, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028567131608724594, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024855418130755424, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.024786634370684624, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022861739620566368, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022829648107290268, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014840544201433659, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014592193067073822, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014509545639157295, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009223484434187412, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.028567131608724594, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.42.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.2045176923274994, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18463772535324097, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.17460346221923828, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15765157341957092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09261965751647949, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08444897085428238, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.11177074909210205, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10319796204566956, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09567952156066895, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.08179189264774323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07815707474946976, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.0565054751932621, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.049232687801122665, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04455673322081566, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04339762032032013, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02839544601738453, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023410551249980927, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02287984825670719, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.02131044864654541, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.020581424236297607, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.01527633797377348, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.01540925819426775, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.013825083151459694, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010687815956771374, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.02839544601738453, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.08457926660776138, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07919758558273315, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07692068815231323, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06981035321950912, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.039433807134628296, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.037165697664022446, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.044690754264593124, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04150786250829697, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03997413441538811, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03534573316574097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03379259631037712, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02264465019106865, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.019811850041151047, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01889204792678356, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01867058500647545, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011329318396747112, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009737947955727577, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009643422439694405, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.00891123991459608, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008821029216051102, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005942917428910732, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005988813936710358, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005662319716066122, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0039657773450016975, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02264465019106865, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07228142768144608, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.0676407665014267, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06510140001773834, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05899951234459877, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.033448439091444016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.031192343682050705, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03897895663976669, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03619451820850372, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.033896367996931076, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.030007854104042053, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.028767896816134453, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019669534638524055, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017244797199964523, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016010141000151634, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.01570086181163788, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009849936701357365, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.00821656733751297, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.008069491013884544, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007523048669099808, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.007373700849711895, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005166572518646717, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005084979347884655, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004804836120456457, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032583391293883324, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.033448439091444016, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.2110619843006134, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19771096110343933, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19369053840637207, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17573317885398865, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09808052331209183, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09324028342962265, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10943694412708282, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10077172517776489, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09934376925230026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08795930445194244, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08412996679544449, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05538338050246239, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.047969523817300797, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.046739913523197174, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04644981399178505, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027623599395155907, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023602042347192764, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023511644452810287, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.02150426059961319, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021436525508761406, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01411389373242855, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013596581295132637, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013621438294649124, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00810181349515915, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027623599395155907, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.16457603871822357, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.13699059188365936, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.12828774750232697, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.1027769148349762, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.07361461222171783, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.06355812400579453, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.08675214648246765, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.07957759499549866, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.07700055837631226, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.05600406229496002, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.05064725875854492, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.044295597821474075, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.038196153938770294, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.035511624068021774, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03485993668437004, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022269833832979202, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01886170729994774, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.018702784553170204, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.015530568547546864, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01524912379682064, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.012223995290696621, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012502441182732582, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01139712892472744, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00910077802836895, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.022269833832979202, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.17434705793857574, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.164499893784523, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.16157248616218567, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14774957299232483, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08178269118070602, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07818188518285751, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.09020968526601791, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08358174562454224, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0827229917049408, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07401129603385925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07103505730628967, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04586413875222206, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.039925262331962585, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03912061080336571, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.038927119225263596, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022904351353645325, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.020080912858247757, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02003156766295433, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018504329025745392, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01847992278635502, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011979809030890465, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012030526995658875, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01171355415135622, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007932834327220917, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022904351353645325, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.2152000218629837, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.2030707448720932, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.199420765042305, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.18241792917251587, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.10084359347820282, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09643708914518356, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.11129061132669449, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.10305100679397583, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.10202264040708542, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.0912809744477272, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.0875583291053772, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05657651647925377, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.049175769090652466, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04818414896726608, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04794755205512047, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02822067216038704, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.024566957727074623, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.02450178749859333, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.022601228207349777, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.022568795830011368, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.014685822650790215, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.014457893557846546, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.014358662068843842, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.009181130677461624, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02822067216038704, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.43.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.20090311765670776, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.18043680489063263, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.1704004853963852, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.15383802354335785, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.09092699736356735, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08248201757669449, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.1104830652475357, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.10138998180627823, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09403102844953537, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07991895079612732, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07643815875053406, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05598701909184456, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04840468615293503, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04375674948096275, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04260685294866562, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028238732367753983, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023018725216388702, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.022498592734336853, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.020889323204755783, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.020174019038677216, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015472901053726673, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015203161165118217, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.014034450985491276, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.010606244206428528, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.028238732367753983, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07611773908138275, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07120965421199799, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06910351663827896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06266383081674576, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03545893356204033, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.03337761387228966, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.040240328758955, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03737136721611023, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03594806790351868, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0317491851747036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03034789115190506, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.020404795184731483, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.017860203981399536, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016996901482343674, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01679115742444992, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.010211396962404251, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.008795167319476604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.008705825544893742, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008051366545259953, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007965835742652416, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005374481901526451, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005460954271256924, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.00511274766176939, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003674785140901804, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0317491851747036, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.07387922704219818, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06910178065299988, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06671112030744553, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.06049094721674919, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03413217142224312, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.031961265951395035, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03944704681634903, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.0365961492061615, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03459152206778526, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03059203177690506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.02928164228796959, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.019878897815942764, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.017401572316884995, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.016315709799528122, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.016054565086960793, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009937074035406113, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.008362798020243645, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.00823865458369255, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007639951538294554, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0075168306939303875, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.005198179744184017, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.005119092762470245, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004869390744715929, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0032608259934931993, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.03059203177690506, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.20998218655586243, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.19677318632602692, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19270436465740204, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17499946057796478, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09760095179080963, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09275493025779724, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10878610610961914, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10028360784053802, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.0988476574420929, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08764318376779556, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0837993249297142, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05505916476249695, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.047777675092220306, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04655345529317856, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04626728966832161, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027470577508211136, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02354380488395691, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023443125188350677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021457122638821602, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021390564739704132, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014058869332075119, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013598553836345673, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013587726280093193, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008156105875968933, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027470577508211136, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.14994126558303833, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.11782211065292358, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.10703777521848679, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07974743843078613, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0665765106678009, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.05456672981381416, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07989604771137238, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.0730152279138565, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.0701901912689209, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.046110108494758606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.041656557470560074, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.041073285043239594, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03552607446908951, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03266096115112305, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.03196293115615845, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.02098427154123783, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01819552853703499, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.01802973449230194, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.014664770103991032, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.01435561291873455, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01217355765402317, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.01308785006403923, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.011348184198141098, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.010475954972207546, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.03266096115112305, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.1725219339132309, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.16283516585826874, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15989671647548676, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.14620627462863922, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08091283589601517, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07736299932003021, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08925052732229233, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.08267948776483536, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08185190707445145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07323946803808212, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07026154547929764, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.045362651348114014, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.039501965045928955, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03870020806789398, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.0385146327316761, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022654365748167038, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.019871562719345093, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.019818726927042007, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.018312934786081314, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01828993856906891, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011850406415760517, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011909237131476402, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011591906659305096, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007862490601837635, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.022654365748167038, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.20346160233020782, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.19201084971427917, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.18859392404556274, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.17249161005020142, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.09532706439495087, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.09115450829267502, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.10522452741861343, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09740670025348663, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.09643331170082092, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08630542457103729, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.08281015604734421, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.05347582325339317, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.046486686915159225, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04555026814341545, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04532884433865547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.026684006676077843, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.02323140576481819, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.023169049993157387, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.021376406773924828, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.02134711481630802, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.01389474980533123, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01367761380970478, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.013586390763521194, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008701333776116371, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.026684006676077843, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.44.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.19692490994930267, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.17587435245513916, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.16598674654960632, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.14918000996112823, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.08920954912900925, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.08067283034324646, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.10785258561372757, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.09913760423660278, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.09231270104646683, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.07774988561868668, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.07418638467788696, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.05487509071826935, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04749206453561783, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.04311065375804901, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.04201939329504967, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027650650590658188, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.023006334900856018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.02252434566617012, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.0208420492708683, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.020172901451587677, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.015168013982474804, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.015562498942017555, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.01379475649446249, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.011350683867931366, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.027650650590658188, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.0830991268157959, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.0777711272239685, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07563277333974838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.0686563029885292, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03871805965900421, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.036509644240140915, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.0437883660197258, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.04062662646174431, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03924141079187393, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.0346958190202713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03315727040171623, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.022193146869540215, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0193951353430748, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01854669116437435, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.018342141062021255, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01110091432929039, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009567081928253174, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009481064043939114, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.0087558189406991, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008674033917486668, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005826793611049652, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.005877829156816006, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.0055666035041213036, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.003909486345946789, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03315727040171623, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.069761723279953, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06526995450258255, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06286565214395523, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.05693211033940315, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03224973753094673, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.030116930603981018, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03750436007976532, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03473098203539848, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03269181773066521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.028896518051624298, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027650415897369385, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018902450799942017, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016517814248800278, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015422512777149677, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.015148631297051907, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009456136263906956, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007910544984042645, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007778072264045477, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007232887204736471, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0070941937156021595, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004956473130732775, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.00487317843362689, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.004618731793016195, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0031089261174201965, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03269181773066521, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21022221446037292, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.1969868391752243, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19294892251491547, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.174876406788826, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09761443734169006, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09276226907968521, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.10860341787338257, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.1002214178442955, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.09883128851652145, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08744781464338303, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08355197310447693, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05496671050786972, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04771135374903679, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.04651292413473129, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04623629152774811, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027399683371186256, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.023526404052972794, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023447036743164062, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021422777324914932, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021369829773902893, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014052473939955235, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013610009104013443, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01359754428267479, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.00821653287857771, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.027399683371186256, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.13267692923545837, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.10249549150466919, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.0912078469991684, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.07627248018980026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.05865395441651344, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.045529913157224655, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.07263151556253433, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.06637148559093475, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.06203242763876915, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.04235876724123955, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03995952755212784, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.037367455661296844, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03233880549669266, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.028933294117450714, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.028086811304092407, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.019057622179389, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.016344938427209854, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.016090774908661842, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.013582568615674973, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.013111728243529797, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01109932642430067, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.012110527604818344, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.01013006828725338, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.009806429035961628, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.03233880549669266, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16509149968624115, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.15575245022773743, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.15299586951732635, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.1398935317993164, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07738026231527328, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.073992058634758, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08542825281620026, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0790533572435379, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.07827512919902802, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.07004965096712112, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06724584102630615, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04342235252261162, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.037785064429044724, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.037023819983005524, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03684299811720848, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.021708402782678604, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.019046787172555923, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.018999269232153893, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01756317913532257, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01753983087837696, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011411227285861969, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01146843284368515, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.01116451807320118, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0076456451788544655, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.021708402782678604, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.18976633250713348, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.17908962070941925, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.175895556807518, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.16086120903491974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.08884403854608536, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.0849657654762268, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.0980340912938118, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.09076977521181107, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.08987365663051605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.08042287081480026, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.07717090100049973, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04982456937432289, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.043306782841682434, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.04244454950094223, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.04223581776022911, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024852823466062546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.021638022735714912, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.021580277010798454, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.019906429573893547, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.019878869876265526, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.012913529761135578, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.012722400948405266, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.012627607211470604, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.008070003241300583, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.024852823466062546, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.45.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.1698928326368332, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.15068231523036957, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.1417902559041977, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.12733155488967896, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.07699420303106308, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.06923617422580719, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.09382417052984238, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.08583122491836548, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.07984738051891327, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.0668163076043129, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.0637124702334404, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.04775026813149452, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.04145696014165878, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.037505172193050385, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.03652539849281311, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.024242226034402847, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.020514022558927536, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.020088937133550644, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.01865234039723873, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.018064487725496292, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.013600545935332775, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.014451670460402966, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.012395183555781841, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.01113075576722622, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.024242226034402847, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07934103161096573, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.07433142513036728, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.07221928238868713, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06554437428712845, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.036942727863788605, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.034826092422008514, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.04184206575155258, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03887147456407547, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.037439487874507904, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03311942145228386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03163684159517288, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.021196097135543823, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.018546387553215027, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.01769227348268032, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.01748836599290371, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.01059677079319954, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.009126017801463604, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.009036961942911148, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.008354723453521729, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.008269217796623707, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.005568257998675108, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.00561563391238451, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.005309650208801031, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0037304279394447803, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.03311942145228386, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.06859797984361649, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.06421221047639847, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.06180466711521149, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.056045472621917725, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.03173435106873512, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.02962157130241394, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.03709571808576584, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.03430180996656418, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03215276449918747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02848951891064644, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.027321547269821167, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.018697112798690796, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.016355173662304878, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.015182985924184322, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.014898347668349743, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.009354658424854279, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.007789905648678541, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.007657443173229694, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.007129545323550701, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.006993765011429787, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004906358662992716, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004815062507987022, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.0045681907795369625, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.003077578730881214, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.03215276449918747, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.21409724652767181, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.20056407153606415, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.19651171565055847, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.17810465395450592, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.09956497699022293, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.09466293454170227, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.11097246408462524, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.10217872262001038, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.10085202753543854, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.08921528607606888, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.08521375060081482, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.05617428198456764, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04862545058131218, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.047444432973861694, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.04717734456062317, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028025589883327484, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.024016721174120903, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.023932015523314476, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.021871909499168396, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.021818915382027626, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.014356007799506187, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.013885331340134144, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.013888717629015446, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.008388757705688477, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.028025589883327484, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.11828720569610596, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.09090115875005722, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.08057531714439392, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.06863675266504288, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.0517209954559803, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.04010302200913429, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.06543640792369843, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.059287600219249725, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.05535631999373436, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.037768956273794174, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.03639725595712662, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0335671529173851, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.0289764367043972, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.025679118931293488, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.024848852306604385, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.017394328489899635, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.01479356735944748, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.014559999108314514, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.012499372474849224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.012042464688420296, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.01052771881222725, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.011259245686233044, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.009646402671933174, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.00929346028715372, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.0335671529173851, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.14492681622505188, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1367783546447754, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.13433648645877838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.12284776568412781, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.06787067651748657, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.06491321325302124, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.07500507682561874, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.06935452669858932, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.06864621490240097, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06147943064570427, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.05901612341403961, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.03814612329006195, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03314858302474022, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.03248155862092972, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03232669457793236, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.0190704558044672, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.016730695962905884, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.016688592731952667, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.015438931062817574, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.015417381189763546, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.010049744509160519, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.010101216845214367, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.009828505106270313, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.006770155858248472, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03314858302474022, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.16164281964302063, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.1526004672050476, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.14988932013511658, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.13709323108196259, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.07576369494199753, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.07246948778629303, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.08368167281150818, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.0773976743221283, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0766390785574913, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.06861820071935654, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.06588905304670334, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.04263468459248543, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.03698992729187012, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.0362529419362545, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.03608352318406105, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02131248079240322, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.018686428666114807, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.01863744482398033, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.01724754087626934, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.017222130671143532, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.011274478398263454, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.011296853423118591, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.011036095209419727, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007588542066514492, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.02131248079240322, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.46.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.06498310714960098, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.056608982384204865, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.05245662108063698, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.047039374709129333, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.029751228168606758, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.026026226580142975, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.03789028525352478, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.03407667577266693, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.03104902245104313, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.025395184755325317, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.024383313953876495, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.019450627267360687, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.01667116768658161, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.014728101901710033, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.014244905672967434, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.009971235878765583, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.008351956494152546, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.008139735087752342, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.007587939500808716, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.007299450226128101, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.005775549449026585, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.006233610678464174, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.0051604975014925, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.005029355175793171, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.03104902245104313, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.q_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.06719888001680374, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.06288035213947296, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.06080165505409241, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.05508929491043091, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.031210390850901604, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.02927357330918312, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.03597819432616234, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03324824944138527, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.03162648528814316, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.027948368340730667, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02677960880100727, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.01822960563004017, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01590181514620781, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.014987151138484478, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.014763912186026573, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.009151342324912548, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.007810360752046108, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.007708367891609669, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.007164405658841133, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.007062677759677172, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.0048904139548540115, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.004952943883836269, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.004619094077497721, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.0034103533253073692, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03324824944138527, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.k_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.05790539085865021, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.054119329899549484, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.051880352199077606, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.04692929983139038, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.026700597256422043, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.024795951321721077, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.031474120914936066, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.029135312885046005, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.027080975472927094, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.02390211448073387, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.022918932139873505, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.015846356749534607, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.013880803249776363, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.012781641446053982, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.012515483424067497, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.007939441129565239, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.0065850913524627686, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.006455127149820328, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.006023834925144911, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.0058877416886389256, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.004195567686110735, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.004124914761632681, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.00387923838570714, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.0026808984111994505, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.031474120914936066, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.v_proj", + "numel": 8388608, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17333984375, + "total_bits": 18231296.0, + "err": 0.1791139394044876, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37646484375, + "total_bits": 19935232.0, + "err": 0.16766896843910217, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62646484375, + "total_bits": 22032384.0, + "err": 0.16406257450580597, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.72021484375, + "total_bits": 22818816.0, + "err": 0.14868086576461792, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.22021484375, + "total_bits": 27013120.0, + "err": 0.08322010189294815, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73583984375, + "total_bits": 31338496.0, + "err": 0.07895451039075851, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0316162109375, + "total_bits": 25431040.0, + "err": 0.09326937049627304, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12646484375, + "total_bits": 26226688.0, + "err": 0.08576159179210663, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17333984375, + "total_bits": 26619904.0, + "err": 0.08430420607328415, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53271484375, + "total_bits": 29634560.0, + "err": 0.07449136674404144, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656982421875, + "total_bits": 30676992.0, + "err": 0.0712377205491066, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0316162109375, + "total_bits": 33819648.0, + "err": 0.047241419553756714, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12646484375, + "total_bits": 34615296.0, + "err": 0.04084993153810501, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.22021484375, + "total_bits": 35401728.0, + "err": 0.039711300283670425, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31396484375, + "total_bits": 36188160.0, + "err": 0.03945178538560867, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.023582834750413895, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.22021484375, + "total_bits": 43790336.0, + "err": 0.02014162763953209, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31396484375, + "total_bits": 44576768.0, + "err": 0.020054807886481285, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53271484375, + "total_bits": 46411776.0, + "err": 0.01833731308579445, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70458984375, + "total_bits": 47853568.0, + "err": 0.018267756327986717, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0316162109375, + "total_bits": 50596864.0, + "err": 0.01213375199586153, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12646484375, + "total_bits": 51392512.0, + "err": 0.011758552864193916, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2191162109375, + "total_bits": 52169728.0, + "err": 0.01171102188527584, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12646484375, + "total_bits": 68169728.0, + "err": 0.007199122570455074, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:5b 128g s4", + "bpw": 5.0316162109375, + "total_bits": 42208256.0, + "err": 0.023582834750413895, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.self_attn.o_proj", + "numel": 67108864, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.17205810546875, + "total_bits": 145764352.0, + "err": 0.07482459396123886, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 159395840.0, + "err": 0.05618411675095558, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 176173056.0, + "err": 0.04997078329324722, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.71893310546875, + "total_bits": 182464512.0, + "err": 0.04137500375509262, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03272505849599838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.73455810546875, + "total_bits": 250621952.0, + "err": 0.025231236591935158, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 203426816.0, + "err": 0.040732648223638535, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 209727488.0, + "err": 0.03648922219872475, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.17205810546875, + "total_bits": 212873216.0, + "err": 0.034916508942842484, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.53143310546875, + "total_bits": 236990464.0, + "err": 0.023189784958958626, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656341552734375, + "total_bits": 245372928.0, + "err": 0.02214958332479, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 270535680.0, + "err": 0.02105226367712021, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 276836352.0, + "err": 0.01854429766535759, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.21893310546875, + "total_bits": 283127808.0, + "err": 0.016951080411672592, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.31268310546875, + "total_bits": 289419264.0, + "err": 0.016556784510612488, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 337644544.0, + "err": 0.011347812600433826, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.21893310546875, + "total_bits": 350236672.0, + "err": 0.010717876255512238, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.31268310546875, + "total_bits": 356528128.0, + "err": 0.010631974786520004, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.53143310546875, + "total_bits": 371208192.0, + "err": 0.009372210130095482, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.70330810546875, + "total_bits": 382742528.0, + "err": 0.009195705875754356, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 404753408.0, + "err": 0.007420509587973356, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 411054080.0, + "err": 0.008832384832203388, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.2187957763671875, + "total_bits": 417336320.0, + "err": 0.007019766606390476, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 545271808.0, + "err": 0.007936840876936913, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.21893310546875, + "total_bits": 216018944.0, + "err": 0.03272505849599838, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.gate_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.12241563200950623, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.11529821157455444, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.11316622793674469, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.10315559804439545, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.05738338455557823, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.05480225011706352, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.06350893527269363, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.058704957365989685, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.05805454030632973, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.05183864384889603, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.049763184040784836, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.03254944086074829, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.02836507558822632, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.02776821330189705, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.027628296986222267, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.016307232901453972, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.014862245880067348, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.014829170890152454, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.013810734264552593, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.013792823068797588, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.008943576365709305, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.009774248115718365, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.008756398223340511, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.007462393958121538, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.03254944086074829, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.up_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.171943132267442, + "total_bits": 391720960.00000006, + "err": 0.12315410375595093, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.375068132267442, + "total_bits": 428355584.00000006, + "err": 0.11593123525381088, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.625068132267442, + "total_bits": 473444352.00000006, + "err": 0.11375270783901215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.718818132267442, + "total_bits": 490352640.00000006, + "err": 0.10372777283191681, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.218818132267442, + "total_bits": 580530176.0, + "err": 0.05778438225388527, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.734443132267442, + "total_bits": 673525760.0, + "err": 0.05518873408436775, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312670330668605, + "total_bits": 546704384.0, + "err": 0.06408926099538803, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.125068132267442, + "total_bits": 563621888.0, + "err": 0.05915745720267296, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.171943132267442, + "total_bits": 572076032.0, + "err": 0.0584777295589447, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.531318132267442, + "total_bits": 636891136.0, + "err": 0.052200622856616974, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.656284066133721, + "total_bits": 659429376.0, + "err": 0.05012090876698494, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.03284966200590134, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.125068132267442, + "total_bits": 743976960.0, + "err": 0.028642872348427773, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.218818132267442, + "total_bits": 760885248.0, + "err": 0.02804429456591606, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.312568132267442, + "total_bits": 777793536.0, + "err": 0.02790115401148796, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.03126703306686, + "total_bits": 907414527.9999999, + "err": 0.01651795208454132, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.218818132267442, + "total_bits": 941240320.0, + "err": 0.01513180322945118, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.312568132267442, + "total_bits": 958148608.0, + "err": 0.015095806680619717, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.531318132267442, + "total_bits": 997601280.0, + "err": 0.014076313003897667, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.703193132267442, + "total_bits": 1028599808.0, + "err": 0.01405661553144455, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.03126703306686, + "total_bits": 1087769600.0, + "err": 0.009167821146547794, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.125068132267442, + "total_bits": 1104687104.0, + "err": 0.01009313389658928, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.21876703306686, + "total_bits": 1121586176.0, + "err": 0.008984306827187538, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.125068132267442, + "total_bits": 1465397248.0, + "err": 0.0078432597219944, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "1.0:4b 128g s4", + "bpw": 4.03126703306686, + "total_bits": 727059455.9999999, + "err": 0.03284966200590134, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + }, + { + "key": "model.layers.47.mlp.down_proj", + "numel": 180355072, + "options": [ + { + "desc": "0.05:3b/0.95:2b 32g s4", + "bpw": 2.177508686864099, + "total_bits": 392724736.0, + "err": 0.048938654363155365, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:3b/0.75:2b 32g s4", + "bpw": 2.37518310546875, + "total_bits": 428376320.0, + "err": 0.04161641001701355, + "qparams": { + "group_size": 32, + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.25:4b/0.75:2b 32g s4", + "bpw": 2.62518310546875, + "total_bits": 473465088.0, + "err": 0.03656040132045746, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.032842062413692474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:4b/0.9:3b 32g s4", + "bpw": 3.2240203147710758, + "total_bits": 581468416.0, + "err": 0.022377725690603256, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.2:6b/0.8:3b 32g s4", + "bpw": 3.7182063612827037, + "total_bits": 670597376.0, + "err": 0.01838366873562336, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 3 + ], + "bits_prop": [ + 0.2, + 0.8 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 128g s4", + "bpw": 3.0312957763671875, + "total_bits": 546709568.0, + "err": 0.03286987543106079, + "qparams": { + "group_size": 128, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:3b 32g s4", + "bpw": 3.12518310546875, + "total_bits": 563642624.0, + "err": 0.02815185859799385, + "qparams": { + "group_size": 32, + "bits": [ + 3 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:4b/0.95:3b 32g s4", + "bpw": 3.177508686864099, + "total_bits": 573079808.0, + "err": 0.023559048771858215, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:4b/0.6:3b 32g s4", + "bpw": 3.5263458961664242, + "total_bits": 635994368.0, + "err": 0.01924126036465168, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.6:4b/0.4:3b 64g s4", + "bpw": 3.6614287620367008, + "total_bits": 660357248.0, + "err": 0.018904225900769234, + "qparams": { + "group_size": 64, + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.6, + 0.4 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 128g s4", + "bpw": 4.0312957763671875, + "total_bits": 727064640.0, + "err": 0.017359010875225067, + "qparams": { + "group_size": 128, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:4b 32g s4", + "bpw": 4.12518310546875, + "total_bits": 743997696.0, + "err": 0.014316173270344734, + "qparams": { + "group_size": 32, + "bits": [ + 4 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:5b/0.9:4b 32g s4", + "bpw": 4.224020314771075, + "total_bits": 761823488.0, + "err": 0.011675004847347736, + "qparams": { + "group_size": 32, + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:4b 32g s4", + "bpw": 4.3228575240734015, + "total_bits": 779649280.0000001, + "err": 0.010971072129905224, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:5b 128g s4", + "bpw": 5.0312957763671875, + "total_bits": 907419712.0, + "err": 0.009335082024335861, + "qparams": { + "group_size": 128, + "bits": [ + 5 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:6b/0.9:5b 32g s4", + "bpw": 5.224020314771075, + "total_bits": 942178560.0, + "err": 0.00742729939520359, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.05:8b/0.05:6b/0.9:5b 32g s4", + "bpw": 5.334485431050146, + "total_bits": 962101504.0000001, + "err": 0.007108767982572317, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.05, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.4:6b/0.6:5b 32g s4", + "bpw": 5.526345896166425, + "total_bits": 996704512.0, + "err": 0.0069662523455917835, + "qparams": { + "group_size": 32, + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.4, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.3:6b/0.6:5b 32g s4", + "bpw": 5.724020314771075, + "total_bits": 1032356096.0, + "err": 0.006581627298146486, + "qparams": { + "group_size": 32, + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.3, + 0.6 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 128g s4", + "bpw": 6.0312957763671875, + "total_bits": 1087774784.0, + "err": 0.0061287591233849525, + "qparams": { + "group_size": 128, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:6b 32g s4", + "bpw": 6.12518310546875, + "total_bits": 1104707840.0, + "err": 0.006330959033221006, + "qparams": { + "group_size": 32, + "bits": [ + 6 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + }, + { + "desc": "0.1:8b/0.9:6b 128g s4", + "bpw": 6.228970194971839, + "total_bits": 1123426368.0, + "err": 0.005237530451267958, + "qparams": { + "group_size": 128, + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "scale_bits": 4 + } + }, + { + "desc": "1.0:8b 32g s4", + "bpw": 8.12518310546875, + "total_bits": 1465417984.0, + "err": 0.005506628192961216, + "qparams": { + "group_size": 32, + "bits": [ + 8 + ], + "bits_prop": [ + 1.0 + ], + "scale_bits": 4 + } + } + ], + "best_option": { + "desc": "0.1:4b/0.4:3b/0.5:2b 32g s4", + "bpw": 2.7240203147710758, + "total_bits": 491290880.00000006, + "err": 0.032842062413692474, + "qparams": { + "group_size": 32, + "bits": [ + 4, + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.4, + 0.5 + ], + "scale_bits": 4 + } + } + } + ], + "base_perplexity": 5.990169370962167, + "q_last_module_idx": 98, + "cal_perplexity": 5.9930964021137925 +} \ No newline at end of file