diff --git "a/measurement.json" "b/measurement.json" new file mode 100644--- /dev/null +++ "b/measurement.json" @@ -0,0 +1,93655 @@ +{ + "last_module_idx": 98, + "measurement": { + "lm_head.linear": null, + "model.layers.0.mlp": [ + { + "accuracy": 0.9209367852461965, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9216925846902948, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9475004578891554, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.961847113935571, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9582964087787428, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9619004506813853, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9743443272615734, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.976185867660924, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9792885952874234, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9787323302344272, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9815959444171504, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.98911110743096, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9907907992601395, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.993726771128805, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946070878129256, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9976725627325083, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981322747312094, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.0.self_attn": [ + { + "accuracy": 0.8095817565917969, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.8293359154149106, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.8430727657518888, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9081932431773135, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9109449512080142, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9115303943031713, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9451506828006945, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9455395121323436, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.949626718696795, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9532256314629003, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9547608620242068, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9580533410373487, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9605483193146556, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9637721588737086, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9781279046284524, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9818023505963778, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9856191167705938, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9888958201596612, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962382893029013, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.1.mlp": [ + { + "accuracy": 0.9822843796328494, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.986279225663135, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860756287449285, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9861323958949039, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937413201520318, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944682823199975, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944576346560529, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9953494699377763, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957633881192458, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968357123434544, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997316684181753, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9983015607454275, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985474996072682, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998552644919408, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989710444290387, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989711165820298, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992040250156271, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.1.self_attn": [ + { + "accuracy": 0.984012013987491, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9870063772327021, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9883798133385809, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9929190583919224, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9930141458385869, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932340932519812, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947486472757239, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950226215939773, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9954647947298852, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956392771319339, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965722729500971, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968446967633147, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968513503278557, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9971449863361684, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982395015264812, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985389585949873, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998608751908729, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991705502922598, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9995603045742762, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.10.mlp": [ + { + "accuracy": 0.9582747409218236, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.959552222176602, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9661346140660738, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9682385199948361, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9786439067439029, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9805273052893186, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9833708342752958, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9887686646298358, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9898503540377868, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890874067419454, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905281082579964, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994429233827089, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952729709054294, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969527401814335, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9971593639960414, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99777150507036, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991761388159112, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.10.self_attn": [ + { + "accuracy": 0.9778951045713926, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9791512850083803, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.981385715697941, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9857235503824133, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9887912242036117, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890004906215166, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932558716912019, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936295826184122, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942481427600509, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945825785398483, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944498233105007, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948623454884479, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952789690149458, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957530957303549, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973742091342023, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978848309501221, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981572988786196, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9988991498555008, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9995100188412165, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.11.mlp": [ + { + "accuracy": 0.9590078968750804, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9602087045970716, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9667173059363114, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9687556969492059, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9791960967214484, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9809320521982092, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9837214350700378, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891614506119176, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901533464067861, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9894099870794698, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9907677314783397, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946119918635017, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9954005695487324, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970758784758417, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972628253070932, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978332117592034, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99922498619478, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.11.self_attn": [ + { + "accuracy": 0.9738951952833879, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9760669407091642, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9783449831761812, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9845430537274009, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987058561099203, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873510010932621, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9916380290922365, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9920818962548908, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992783178624354, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9931064004960813, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936245075966182, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940958293645006, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945400612134683, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949816359501136, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969148488813325, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997475586047298, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978241020519483, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985956180056459, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9994122849189138, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.12.mlp": [ + { + "accuracy": 0.9546483284548709, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9558794467072738, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9628897811237135, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9650804651410956, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9768327866729937, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9787584998105702, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9818284103744909, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880141553125883, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891190434757032, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882024950102756, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9897125826070183, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940030237561778, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948804464779402, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996776499442364, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969435173429941, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975637944513246, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991417332111221, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.12.self_attn": [ + { + "accuracy": 0.9728261483343024, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9746557881957606, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9769970900134036, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835245436743686, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9863951614028529, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9865865503486834, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9919009059667587, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9920779654854223, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9926993431229341, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9931692222231313, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932258089906291, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99370928030265, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942535573714658, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947276362463048, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996824011010559, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973660370236949, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978950978501847, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985554591801605, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9994358503700871, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.13.mlp": [ + { + "accuracy": 0.9517399856918737, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9531216558657194, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9606251528388575, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9629214656980414, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.975393229409268, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9775034361764005, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9807137376383731, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9872370377967232, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9884217389320072, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9874111970788554, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890613932358591, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935733903395502, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945431300684026, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965350118122602, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996712572088367, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973736501445896, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990812022435037, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.13.self_attn": [ + { + "accuracy": 0.9730673400979293, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9748106081234781, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9772556612366124, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9834665295324827, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9864040377892946, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9866882876345986, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915963379960311, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9919470289820119, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992570698261261, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992998618828623, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932368891803842, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937093669646665, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942445351104987, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946750312259323, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9967927334732131, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973442648586474, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978099486937648, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985203776312502, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9994072122009177, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.14.mlp": [ + { + "accuracy": 0.9508766249606484, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9523174511758905, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9601816503625167, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9625495734967684, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9750080845857921, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9771723057094374, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.980483425290961, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9870311955088064, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882457397486034, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987227442233186, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.988906885448255, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9934773794130275, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944670231718766, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964811394089147, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996669663018302, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973587736879524, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990681567670483, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.14.self_attn": [ + { + "accuracy": 0.9718791371897647, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9734959445501629, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9758881565771604, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9822679773757332, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9858097402673018, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.986055380419681, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914580356133612, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.991708589227576, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992498381357444, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9929114791907763, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9929810448696739, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.993478540919329, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9939839894834318, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994490343489145, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966787609614824, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972355730439487, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977924278692195, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985137904356969, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9994080037270722, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.15.mlp": [ + { + "accuracy": 0.9510062870226408, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9525556752556249, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9605156842030977, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9628969179956537, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9752119748215926, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9773915708065033, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9806915741217764, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9871173336317665, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9883119307066265, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873332538102803, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890157784286299, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935351347452716, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945197513228968, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964940983213877, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966968290115658, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997378684776394, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990645178446644, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.15.self_attn": [ + { + "accuracy": 0.9687208250949257, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9706122122312847, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9736064703840959, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9803387852091539, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.984394393469158, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9846981308962169, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910816503198523, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914103236637617, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9920699376808969, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925443310486642, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923168828612879, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9928498770061293, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935621532954668, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940899747766947, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964980761471548, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970427584883413, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977288618683815, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984207713094196, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9993860090249463, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.16.mlp": [ + { + "accuracy": 0.9526293371853075, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9540712770662809, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9618122201216848, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9641612297610233, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9758439095396745, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9780152126362449, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9812841148752915, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873872305217543, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9886126071214676, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987633878463193, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9893023842259457, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936829256383997, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946583078095788, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965696346602941, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996768348311123, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974555706507281, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990788213908672, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.16.self_attn": [ + { + "accuracy": 0.971585894885816, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9733304569595739, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9757685488776157, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9823427090519353, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9859001730617724, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860828377698597, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9917271215664712, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9919597388882386, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9924589591590982, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9928189892517892, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9928553433794725, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9934347413088146, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9939652528417738, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944686921019303, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966789056596003, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997246554219409, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978220835328102, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985060688892478, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.999414254715176, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.17.mlp": [ + { + "accuracy": 0.9537878946254128, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9552033198507208, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9626771086140683, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9649468346645957, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9764212714998346, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9785104958634627, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9816940913074895, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9876936878028669, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.98886980430076, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9879264745273089, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895315421255011, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938293024897575, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994771311549764, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966478418362769, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968458494465602, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975162217099416, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991058553323934, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.17.self_attn": [ + { + "accuracy": 0.969792554253026, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9715045690536499, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9735354655667355, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9797590625913519, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.984767253461637, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9850923764078241, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908804156278309, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.991328112388912, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9920700615958163, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992538363525742, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9924250582331106, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9930662238284161, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935029164740914, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9941265602645121, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964104953565096, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970687496426859, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997656635744007, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998402287987502, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9993724507328711, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.18.mlp": [ + { + "accuracy": 0.9542058518058375, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9556334332415932, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.963057279586792, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9653180774889494, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9766104456625486, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9787023726262545, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9818735671670813, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9877602293303138, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889492721934068, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880130338041406, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9896184872639807, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938683007892809, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948134336032366, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966663600582826, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968597892867891, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975264725324354, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991033802691259, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.18.self_attn": [ + { + "accuracy": 0.9674831628799438, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9693350384109899, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9722891136219627, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9785437050618624, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835671045278248, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.984015952599676, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9896845731296038, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9902151446593436, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912697641473067, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.991740597706092, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9920767922150461, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9926930705183431, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932577453161541, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938222946305024, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962616931450995, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969086621544863, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974891059100628, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982888198604709, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9993002480876289, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.19.mlp": [ + { + "accuracy": 0.9537702359651264, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9552444250960099, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9628457834846095, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9651798700031481, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9763120491253702, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9784810778341795, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9817066255368685, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9875725492050773, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9887917379015371, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9878446757793427, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9894992473878359, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937807203907716, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947504412971044, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966127750904936, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968170025631001, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975027130230477, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990956881328633, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.19.self_attn": [ + { + "accuracy": 0.9704961306170413, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9721665131418329, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.974505774284664, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9802221922498, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9849432471551394, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9851591728235546, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9906413994337383, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9909976924720564, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9919790155009219, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923418956367593, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925831611219206, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9931290910432213, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935916128911471, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9941668232020578, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964274042531064, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970723343522925, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975998123225412, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9983711067195001, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9993443156739599, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.2.mlp": [ + { + "accuracy": 0.9664939547839918, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9702020726705852, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9704237235219855, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9729467410790292, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9862693014897799, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9875750400518116, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9875435068419105, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895348039112593, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908436356406463, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9930481291130969, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99405385161701, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963891368947531, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969520864910201, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970734072358984, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982352662635477, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998361970935213, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9987761455735094, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.2.self_attn": [ + { + "accuracy": 0.9649534413689062, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9693382727472406, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9708728413832816, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9832087331696561, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.983425088618931, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9836345691429941, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901941293164304, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904340787937767, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910956338832253, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9916117936372757, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9916519420711618, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992216460798916, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9926678546165165, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932605153635928, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959383595146631, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966063719046744, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974123227753138, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981637879421836, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992783840157484, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.20.mlp": [ + { + "accuracy": 0.9531643327913786, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9547784453944156, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.962515978436721, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9649014786670083, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9760037786082217, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9782133965115798, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9815250525349065, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873878751930437, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.988605860816805, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9876895813565505, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9893597168357748, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.993696748426086, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946789667010307, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965489628283601, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9967707309283709, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974842489157852, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990661536392412, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.20.self_attn": [ + { + "accuracy": 0.9673969526039927, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9692731593784533, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9715466624812076, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9783753423314345, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835527413769772, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9838259392663052, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9899062137854727, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901925741057647, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9909418711536809, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.991528969062002, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9917597974601545, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923057297342702, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9928531976122605, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935085114679838, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9960773940149107, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9967390396877339, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973628746443673, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99814519070481, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992986418993065, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.21.mlp": [ + { + "accuracy": 0.9539666269954882, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9555353616413317, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.963322501433523, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.965712983357279, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9764072581341392, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9785532590590025, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9818768956159291, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9875386231823972, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9887691493097105, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9878754458929363, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895145712714446, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937926248500222, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947522419847941, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965939721778819, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968212354732188, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975456171914151, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990876916992036, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.21.self_attn": [ + { + "accuracy": 0.9634903512502971, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9655538609153346, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.967359018953223, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9732612371444702, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9823703954094335, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9825818538665771, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9896661772539741, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9898028318819246, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908905037139591, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9918149074441508, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9913765356728905, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992162223709257, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9927923381328583, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933169868431593, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9960081663570906, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966728063790422, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974820717777076, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982126910043391, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9993201993699921, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.22.mlp": [ + { + "accuracy": 0.9506591495714689, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9539152383804321, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9606193837366606, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9626792355587608, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9766669430230793, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9790010326787045, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9820611351414731, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987812810822537, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9888745318902167, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.988039670806182, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9896728780708814, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938850293034002, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947465673873299, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965929922304655, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968570557079817, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997522703322925, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990316794106835, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.22.self_attn": [ + { + "accuracy": 0.9658237131018388, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9676570108062342, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9699371331616452, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9772364638353649, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9824069220768777, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9827825442740792, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989470047386069, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9899830677007374, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9906475755729174, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912827101192976, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912146016171104, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9918187621392702, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923560736994994, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9930483879227388, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957323411577627, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965160120474664, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972521309789858, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981083099387194, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992680192777985, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.23.mlp": [ + { + "accuracy": 0.9564045291197927, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9578608588168496, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9647049590160972, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9668211591871161, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9774752055343828, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9797298908233643, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9826437156451376, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880796388575905, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9893116723549994, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9883658462449124, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9900367863868412, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940101065133747, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.995002050933085, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966950957712374, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969253796888026, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9975404249210107, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.999104279926733, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.23.self_attn": [ + { + "accuracy": 0.9654473662376404, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9671347517716258, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9698409845954493, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9760625111429315, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9824173105390448, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.982422577707391, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9897484826414209, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989948993450717, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908929953449651, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915950031656968, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910617266830645, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9917015833289999, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992614212004762, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932669314898943, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959937412487833, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966705837532094, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973166592811283, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982201909940494, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992938194619981, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.24.mlp": [ + { + "accuracy": 0.9584145294992548, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9597537203838951, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9664071359132466, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.968459195212314, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9787235620774721, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9806157193685833, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835051831446195, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9888166228407308, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9898780717661506, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890829073755365, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905283560878352, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944071103083459, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952603896197519, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996946066029762, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9971390699869708, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977573070086931, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991739832453037, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.24.self_attn": [ + { + "accuracy": 0.9659694903775266, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.968355793701975, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9707993676787928, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9773177190830833, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9830018203509482, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9832735924344314, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891431684556761, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9894161804726249, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901867618686274, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990740650578549, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.991537625852384, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9921383049927259, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992591626550022, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932851414931448, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959220164700558, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996590790779967, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972255443663973, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9980414490166464, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.999242400211331, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.25.mlp": [ + { + "accuracy": 0.956134767908799, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9575535874617727, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.964612813372361, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9668002410938865, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9774776116797799, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9794957512303403, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9826033883973172, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881655636586641, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9892999208287189, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9884352229143444, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989968581419242, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940733176312948, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949813532201868, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9967712322740179, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969646118973431, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9976332711153909, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991273478065666, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.25.self_attn": [ + { + "accuracy": 0.967774162166997, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9692237596762808, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9715511579262582, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9773208436213041, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9833350416861082, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9837353433433332, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895710592207155, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9900874589618883, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9909268029426274, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9911391389997382, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915631993820793, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923875167181617, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925926053210309, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933586740180066, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9958679558415162, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966332657556785, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972169063985348, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9980985486977979, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992327994519943, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.26.mlp": [ + { + "accuracy": 0.9538662872816387, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9553741216659546, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9628392834412425, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.965193930425142, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9760815218875283, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9783380768801037, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9816274799798664, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987356676082862, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9886306383107838, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9876857337198759, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989387035369873, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936824091955235, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994677507955777, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965371932638319, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9967610687017441, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974649723031019, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990567292429899, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.26.self_attn": [ + { + "accuracy": 0.9692593279637789, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9704330375320033, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9728353933284157, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.978497177362442, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.984465562983563, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.984581028160296, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990325781075578, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904334490236483, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9906633347272873, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914013790456873, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9920879665173983, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925933213610398, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9929232589508358, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.993389652747857, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9960928484797478, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964884519577026, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973810311210783, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978202373573655, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992877936578894, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.27.mlp": [ + { + "accuracy": 0.9511566507188898, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9527325190995869, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9602148846576088, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9626409944735075, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9748556331584328, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9775325216745075, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.980946056152645, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9867360042898279, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881264120340347, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9871930417261625, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9888067355281428, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933510940325888, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944085651322415, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962312520334595, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996617855994325, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973521495335981, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998960828212531, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.27.self_attn": [ + { + "accuracy": 0.9624809936473244, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.964356497714394, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.967289758355994, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9739676259065929, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9805401532273543, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9808949263472306, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9871164105440441, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9874586541401712, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9887410637579466, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9893641291480315, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901576489210129, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910492787235662, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915464583196139, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9922538627135126, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9953348958178571, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9961212995020967, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966055174407206, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9979204843310934, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.999108919089562, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.28.mlp": [ + { + "accuracy": 0.9522810515604521, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9539801258789866, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9615308234566137, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.964053279475162, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9753227139774122, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.977617903759605, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9809791194765192, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9867842542497736, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881341943615362, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987336374427143, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890652887131038, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935155091317076, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945292684592699, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963792240933368, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996699039873324, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974417090415955, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990089549437949, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.28.self_attn": [ + { + "accuracy": 0.9635172580417833, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9654001875927574, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9687975613694442, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9754765535655775, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9810703851674732, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9813413180802998, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9875922022681487, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9878315219753667, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890671090075844, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989714494661281, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905003701385698, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912879074874678, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9917659767364201, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925419507842315, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9953625966843805, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962357980640311, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996569562899439, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9979021647258809, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990789155034643, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.29.mlp": [ + { + "accuracy": 0.9505902152312429, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9522964923005355, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9601895369981465, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9628062122746518, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.974491862874282, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9768365668623071, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9803374127337807, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9864218250701302, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9877805035365256, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9869133681058884, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9886921785379711, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933007296762968, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9943452524511438, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962955419170229, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965860286825582, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973495126162705, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989920231072527, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.29.self_attn": [ + { + "accuracy": 0.9617593853097213, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9635173239206013, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9664656394406369, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9733445707120394, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9797798097133636, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9802199915835732, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9866359202485335, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9871147792590292, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882261157035828, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889597492782694, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989561876184062, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9903867181978727, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908832735136935, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914978138710323, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949363885741485, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956652510323023, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963154083019808, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9976766583950896, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990129159077218, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.3.mlp": [ + { + "accuracy": 0.9524648660107663, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9552179010290849, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9583764546795895, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9636302182548925, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.97671976214961, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9788468621279064, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9813789191998934, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9841716728712383, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.986177733069972, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880836206047159, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9897691176125878, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938690027124003, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948362123809362, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956808007861438, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997076712156597, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997876299447135, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984995912956564, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.3.self_attn": [ + { + "accuracy": 0.949880411750392, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9556175470352173, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9580147893805253, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9751775829415572, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9763784800705156, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9766060160963159, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9857912879241141, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860358975435558, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9870079285220096, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9877863338119105, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880747983330175, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9888706505298615, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989455604239514, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9902923444384023, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942099773570111, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9951371307435789, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962852216864887, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973190685636119, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989978776950585, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.30.mlp": [ + { + "accuracy": 0.9492168332401075, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9508847437406841, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9589345768878335, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9616355174466183, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9738561134589346, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9761817376864584, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.979784408682271, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860716509191614, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9874437921925595, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9866229232988859, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9884050990405836, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9931699131664476, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942082880358947, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996218200184797, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965358580413618, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973023341674554, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989719958485741, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.30.self_attn": [ + { + "accuracy": 0.9612776354739541, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9630197443460163, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9666027646315725, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9739803907118345, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9794379676643171, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9800241213095816, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9866884476260135, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873807594964379, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9886448100993508, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9890660510251397, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895681268290469, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904849027332506, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910062433857667, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99185627071481, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949752527632212, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957641206289592, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963117939861197, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977772186853384, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990280784274402, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.31.mlp": [ + { + "accuracy": 0.9470478484505102, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9487481023135939, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9571388865772047, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9600016725690741, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9727445087934795, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9751300780396712, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9789282601130637, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9854026546603755, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9868313215280834, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860711395740509, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9879026350222135, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9928914614413914, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9939606856358679, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9960352264736828, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964054587640261, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972083690135103, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989082597588238, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.31.self_attn": [ + { + "accuracy": 0.9602651031393754, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.96151413101899, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9667026055486578, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9730811997463829, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9806503935864097, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9811428879436693, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9878801899520975, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9883861290781122, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891158499215779, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9896346942374581, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904842478664297, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910784779410613, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9918773770332336, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9924781495019009, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9955234907959637, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962021210475972, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968495298373071, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998011157504822, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991763046893635, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.32.mlp": [ + { + "accuracy": 0.9434107228329307, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9452058829759297, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9542425274848938, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9573136003393876, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9708274916598671, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9733615737212332, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9774687917608964, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9844676350292406, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9859982854441592, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9850763192302302, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9870419070908898, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923819452524185, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935286860716971, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957868923482142, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9961304954792324, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969937868024173, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9988405543722605, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.32.self_attn": [ + { + "accuracy": 0.9529490000323245, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9611011272982547, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9655300472912035, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9720716570553026, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9788395856556139, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9792770664942893, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9857928627415707, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.986303514555881, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9876133642698589, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.988363390690402, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891276241917359, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9902214407920837, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9906531572341919, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915917586339148, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946538276578251, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956046374220597, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9958536252379417, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978956995826018, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998903427469103, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.33.mlp": [ + { + "accuracy": 0.9372231081912392, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9391827144120869, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9491933364617198, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9527040594502499, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9675806980384023, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9704529737171373, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9750276192238456, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9825532514798013, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9842853671626041, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.98341955166114, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9856184407284385, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915367309984408, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9928172285619535, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952434301376343, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957241721843418, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9966880730108211, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998684610975416, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.33.self_attn": [ + { + "accuracy": 0.9571199605339452, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9589390754699707, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9635187167870372, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9701584420706096, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9774142284142343, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9782524846102062, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9850623105701647, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860152617881173, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880023833952452, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9888073574555548, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9892766656059968, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9899388920319708, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990771369714486, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9916646166851646, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950570079841112, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957783371210098, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965052949754816, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978516823367068, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990531874722556, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.34.mlp": [ + { + "accuracy": 0.9349029817079243, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.936878900778921, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.946940447154798, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9505403606515181, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9663997857194198, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9693270733481959, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9739586036456259, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9819156182439703, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9837163796550349, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9828065539661207, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9850888095403972, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912196821288058, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925557884730791, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950830940353242, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9955592673075827, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99652351831135, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9986384994301357, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.34.self_attn": [ + { + "accuracy": 0.9534250686043187, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9558798921735663, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9597706449659247, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9648900345752114, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9754571616649628, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9758285613436448, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9855349801088634, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860251655704096, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880175245435614, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9885022059867257, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9880352420242209, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891302287578583, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9907791536105307, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9917876963552675, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950125327235774, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959073831376276, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963207048805136, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9979205745223322, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989973275285018, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.35.mlp": [ + { + "accuracy": 0.9323679334238956, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9344196131354884, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.944813333059612, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9485824045382047, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9650385944466842, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9681199632192913, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9729463991365934, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9811924068551314, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9830459622960341, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9821282421287737, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9844901702905956, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990861210383867, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9922508220923575, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948774238950328, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9953711362261521, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963769230403399, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985650369015179, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.35.self_attn": [ + { + "accuracy": 0.948815173224399, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9543436608816448, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9582047368350782, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9667900518367165, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9750320644755113, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9756691000963512, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9844526805375752, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9853001895703768, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9862653518977919, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9870949779686174, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987778611089054, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9885829569477784, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895275148906206, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905597881266945, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9943257715356978, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952357183945807, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959837468831163, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974550763635259, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989198459998557, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.36.mlp": [ + { + "accuracy": 0.9284382619355854, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9305734948108071, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9413964560157374, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9452961808756778, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.963064055693777, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.966188299028497, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9712477420505724, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.980195795234881, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9821247991762663, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9811595866554662, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835846188821291, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904088856358277, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9918115115479419, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946620840775339, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9951533561474398, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962110593914986, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985326420711843, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.36.self_attn": [ + { + "accuracy": 0.9519368221885279, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9553584268218592, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9605128702364469, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9669392799076281, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9746539420203159, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9751815780213005, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835441332114371, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9846044512171495, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9864980170601293, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9874322155588552, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9875177490083795, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889132929475684, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9897603925905729, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908844381570816, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9943587313357153, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9955358524855814, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957143056549524, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9976865410020477, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9988194220748386, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.37.mlp": [ + { + "accuracy": 0.9259896780315199, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.928107500076294, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9390969652878611, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9429060910877428, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9617436089013752, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9649857219896818, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9701095254797685, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9796652433119322, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9816175228671024, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.980444544240048, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9829841836502677, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9900356123321935, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915169563732649, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945031365281657, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949559043896826, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959985088360938, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984915316300956, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.37.self_attn": [ + { + "accuracy": 0.9529527112057334, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9558902320108915, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9603530764579773, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.967339835668865, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9756359175631875, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9761939064452523, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.984303996751183, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9851436128741816, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9867577960616664, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873524080765875, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882730731838628, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989202071177332, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9902300073912269, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912476037677965, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944862193967167, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956369443159354, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957875185891202, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9976050700796278, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9988923209082139, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.38.mlp": [ + { + "accuracy": 0.9231204672863609, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9252942173104537, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9363519267032021, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9401462391803139, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9603061581912794, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9635660554233351, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9687552169749611, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9790046246428239, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9809954307581249, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9797442931877939, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9823116126813387, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9896861181447381, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9911888147655287, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9943516972033601, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947764006884474, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9958372457247031, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984714380024295, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.38.self_attn": [ + { + "accuracy": 0.9569349571278221, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9597458274740922, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9638022560822337, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9701944620985734, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9782743814744448, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9787893844278235, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9861382167590292, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9868882447481155, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881918751879742, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9887443773056331, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989333481380814, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9902077442721317, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990894814855174, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9917448390471307, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948529012893376, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9958756173911848, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9961991004253689, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977790348623928, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989804762758707, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.39.mlp": [ + { + "accuracy": 0.9199025379984003, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9221821521457872, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9333425446560508, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9371427423075626, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9586282184249476, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9620032467340168, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9672389469648662, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9781299468718077, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9802108312907972, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9788940172446402, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9815651052876523, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9892574399709702, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9908149375727302, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9941269069125778, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945598656409665, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956154925258536, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984137548231765, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.39.self_attn": [ + { + "accuracy": 0.9533088175874007, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9557997678455553, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9612096171630056, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9680702403972024, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9762755820625707, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9771708814721358, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9855245147880755, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9860990172938296, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9877217733546307, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881523102521896, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9884042104608134, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9894918368050927, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904931728777132, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914781741405788, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946856459504679, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957283300004507, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9961604593615783, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977064364050564, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989439173552551, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.4.mlp": [ + { + "accuracy": 0.9136292181516948, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9300547524502403, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9336520119717246, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9350700629384894, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.979317224339435, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9838032267595592, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.985429186570017, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9894292621236098, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9896632189813414, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987815512638343, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932288056925723, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947847692590011, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9958438496840628, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963316176282732, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969322528494032, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997312519895403, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977467295370603, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.4.self_attn": [ + { + "accuracy": 0.9412579661921451, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9458028580013075, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9492916307951275, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9702703450855455, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9713507796588697, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9716116158585799, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9837045967578888, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9839811230960646, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9850798421784451, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9859962933941891, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9855157372198606, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.986483002963819, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9872573880772841, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882801026105881, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9929992525201095, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9941206746979764, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9957350476791984, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968815930187702, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998859412379955, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.40.mlp": [ + { + "accuracy": 0.917879863789207, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9201503364663375, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9312575741818077, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9350297952953138, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9575174356761732, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9609963329214799, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9662013681311357, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9775738904350683, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9797074637914959, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9783313368496142, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9810625926444405, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889704239995856, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905647141368765, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9939763938125811, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944075024441669, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9954466208031303, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9983741397920408, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.40.self_attn": [ + { + "accuracy": 0.9554689526557922, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9580207561191759, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9637637389333624, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.969393109020434, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.977567594302328, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9782328527224692, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9844686530138317, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9855765276833585, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9870961724143279, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881227322314915, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989020699733182, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990067089858808, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990825249960548, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9918760202432934, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994822538212726, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959283094657095, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959522736699957, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977468444328559, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989133183109132, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.41.mlp": [ + { + "accuracy": 0.9159560140810514, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9181979580929405, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9292339525724712, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9329147527092382, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9564648709799114, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9600628049750077, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.965228469748246, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9771233201026917, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.979285749949907, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9777724099786658, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9805887702264284, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9886774938357504, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9903289699240735, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938502315627901, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942468130274823, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952715556872519, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9983516451167432, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.41.self_attn": [ + { + "accuracy": 0.9580130482974806, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9602270502793162, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9656803545198942, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9711482117050573, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9780243368525254, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9787962891553578, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.985087586076636, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.985950192338542, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9879370206280759, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9885978902641096, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9892797846543161, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990427691685526, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9909622645691821, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9921472260826513, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949449387035871, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9960973639237253, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9960414763343962, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978653335649716, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989503164236483, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.42.mlp": [ + { + "accuracy": 0.9133672463266473, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9157211153130782, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9268148321854441, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9304983490391782, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9551126831456235, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.958877444267273, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9640376160019323, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9763979049105393, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9786507016734073, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9770440584734866, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9799914187506625, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882864944244686, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9900212115363071, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936307817697525, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940374540655237, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950594219722246, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.998285778651112, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.42.self_attn": [ + { + "accuracy": 0.9570518857554385, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9599873203980296, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9647957274788305, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9693878449891743, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9780263006687164, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9792489613357344, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9842288917616794, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9855410644882604, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9874274503243597, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881160674910796, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9895370147730175, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9907959817271483, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910514040997154, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9922835638648585, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949269047693202, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9961070418357849, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.995838647218127, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978235060055005, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9989190038881803, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.43.mlp": [ + { + "accuracy": 0.9123914053565577, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9147738594757884, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9257578410600361, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9294503864489103, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9545612021496421, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9584162266630876, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9635435907464278, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9760399241196481, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.978319547678295, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9767650444256631, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9797457393847013, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9881460180408076, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9898955555338609, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935321835310835, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9939625702406231, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949673249533302, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982521226139445, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.43.self_attn": [ + { + "accuracy": 0.9518083867273832, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.954839405260588, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.958970568682018, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9656020685246116, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9748393955983614, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9750693612977078, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9827324559814051, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9832860341197566, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9850343340321591, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9859244682286915, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873890657173959, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9886550989590193, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9893472265256079, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904636262278808, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9940290721623521, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952583348280505, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952794882027727, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9972738717731676, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9987550968990514, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.44.mlp": [ + { + "accuracy": 0.9075526438261333, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9100635177210757, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9216114031641107, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9255343550129941, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9519815884138408, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9562546893170005, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9617259753377814, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9744820484989568, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9770877753433428, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9754710495471954, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9786562998043863, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987424005803309, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9892711811944058, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9930331063897986, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936120117965498, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946751222014427, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9980725324467609, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.44.self_attn": [ + { + "accuracy": 0.9500975295117027, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.952764586398476, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9580339318827579, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9654584777982611, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9732205491316946, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9740462005138397, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9804822880970804, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9814378744677493, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9842427482730464, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9848520975363882, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9867817571288661, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882877759243313, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882935677704058, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9898222182926378, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933881783171704, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949454794588842, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994649832185946, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970965863842713, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985693487289705, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.45.mlp": [ + { + "accuracy": 0.9107490338777241, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9132715714605231, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9244021302775333, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9283104758513601, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9535152347464311, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9576300664951927, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.96284629482972, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9751900748202675, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9776322873015153, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9761033340504295, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9792667580278296, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9877451085730603, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989631786942482, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932041434865249, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937674830618658, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948137316264605, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981279482966975, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.45.self_attn": [ + { + "accuracy": 0.9475962927466944, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9522468196718317, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9567521183114303, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9635128159272044, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9731502281992059, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9737440975088822, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.981546590202733, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9821909226869282, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9843927166963878, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9852186017914822, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9869331144972852, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882718183492359, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9888376460263604, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9902144482261256, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937395676970482, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950810090491646, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9950481504201889, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.997139371735485, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9986934162872402, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.46.mlp": [ + { + "accuracy": 0.9195449289522672, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9218941926956177, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.931291611571061, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9345833878768117, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9576359799033717, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9619307486634505, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9662260162202936, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9767940624764091, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9790632960043455, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9773172438144684, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9806745617013228, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.987770109584457, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9900467340883455, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9918373352602908, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994235943414663, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9951665793594561, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978967481537869, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.46.self_attn": [ + { + "accuracy": 0.9412224794688978, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9470525729028803, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9530984947555944, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.961548111940685, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9723356899462248, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9733183070232994, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9810121137844888, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9820539370963448, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835122070814434, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9848021425698933, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9863830820510262, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9875407509113613, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882244922612843, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.989335937719596, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9934534348155323, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946765672219428, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.995076513604114, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969029567743602, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9986512497263519, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.47.mlp": [ + { + "accuracy": 0.929541970554151, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9320111274719238, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9382916626177336, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9408541541350515, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9638244854776483, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9671266863220617, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9700850028740732, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9802930339386589, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9821103732836874, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9813205725268314, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9838349803497917, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9903060621336887, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.991768080937235, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942860571961654, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994980479149442, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956424083364638, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9980024916953162, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.47.self_attn": [ + { + "accuracy": 0.9714438570173163, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9735515368612189, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9769761703516308, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9820295321313959, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9863589280530026, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9867232630127355, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9903595494596582, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9907963236695841, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9916005824741564, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9920094930811932, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9931471841899973, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937011387787367, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938378973226798, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9943416275476155, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964894832749116, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9970561226731852, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973032947041487, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984232587249655, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992445095589286, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.5.mlp": [ + { + "accuracy": 0.97498346316187, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9759407388536554, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9793144276267604, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9813408396745983, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9876344815680855, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9882038102338189, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9901209442239058, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925540544484791, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9925757285795713, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9935924026526903, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9943081830677233, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959370470360706, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9971684868398466, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968322458627977, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982878616765926, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9987919218838215, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9991854056715965, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.5.self_attn": [ + { + "accuracy": 0.9691659274854159, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9710551594433031, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9731571360638267, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9828203728324488, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9848189024548781, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9849657134005898, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912665227526113, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9914249762108451, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9919999398683247, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9924778938293457, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992372350473153, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.992861533635541, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933030503360849, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938298644203889, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996365664820922, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969194390271839, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977372227922866, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9983683597100409, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9993968517764619, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.6.mlp": [ + { + "accuracy": 0.973625333685624, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9744518822745273, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9774971400436602, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9790176294351879, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9878585777784649, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9887179183332544, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.990472779462212, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9929496504758534, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932413532545692, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9938299271621203, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947087623571095, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9964178906459558, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9968381375074387, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974298935971762, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9983565606372921, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9987937941362983, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992620467061275, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.6.self_attn": [ + { + "accuracy": 0.97769117982764, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9794903366189254, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9815567151496285, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873756797690141, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889814414476094, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891709755910071, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9933767067758661, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936246871948242, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942000143622097, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945182168954297, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9945019303183806, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948805727456745, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9953228643066004, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956886446789691, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973967851776826, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978379669942354, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982728669910055, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9988388021133447, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9995380027317687, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.7.mlp": [ + { + "accuracy": 0.9725639098568967, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9734780317858646, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.977191697610052, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9787349481331674, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9861050028549997, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9873894492262288, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9891652206056997, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9923187666817715, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9931313591568094, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9928416194100129, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9937739466366015, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9963206366488808, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9969247016859683, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978200652097401, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981581639302405, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985737425128096, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9993658064325389, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.7.self_attn": [ + { + "accuracy": 0.9807688437010113, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9819557745205728, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9840097097974074, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9885734492226651, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9903250524872228, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905098449242743, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9942226284428647, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944476504859171, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9949515818765289, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952176393646943, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9951819255948067, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9955619055973856, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9959129714652112, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9962860694841335, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9977420886096201, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981293313597378, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9984795212941734, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.999012549750899, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9995944866616475, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.8.mlp": [ + { + "accuracy": 0.9653974081340589, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.966524246491884, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9719138772864091, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9737422309423748, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.982402829747451, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9840187242156581, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9863266286097074, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9905636773297661, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9915031751519755, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9909942063846087, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9922226275268354, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.995389183493037, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9961085495980162, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9974149371448316, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9976517773772541, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9981932587137348, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992806809513193, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.8.self_attn": [ + { + "accuracy": 0.982061870788273, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9835211117016641, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9851918189149154, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9893222529637186, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910888867942911, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9912592117723665, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9943680590704868, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9946045534391152, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948770756784239, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952216328758943, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9955556549524006, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9958489862711806, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.996189291932081, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9965199961474067, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978856663954886, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982407914572641, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9985198710898036, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9990417075000311, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9996024487344058, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.9.mlp": [ + { + "accuracy": 0.9604744346518266, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 474871952, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9616875962207192, + "down_proj": { + "bits": [ + 6, + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.2, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64, + "6": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 492566672, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.968375635774512, + "down_proj": { + "bits": [ + 5, + 3 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "3": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 549465856, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9704993367195129, + "down_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 616640256, + "up_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.3, + 0.7 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9798686253397089, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 128, + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 694777456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9816351532936096, + "down_proj": { + "bits": [ + 8, + 4, + 3 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "3": 32, + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 714362112, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9845180197765953, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 768025456, + "up_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "3": 32, + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9894437656590813, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 878169328, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9904782395613821, + "down_proj": { + "bits": [ + 8, + 4 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "4": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 891112704, + "up_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9897052441772661, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 128, + "5": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 903509616, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9910662111483122, + "down_proj": { + "bits": [ + 8, + 5, + 4 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "4": 32, + "5": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 923094272, + "up_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9947441251654374, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 128, + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1112241776, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9955407545754784, + "down_proj": { + "bits": [ + 8, + 6, + 5 + ], + "bits_prop": [ + 0.05, + 0.1, + 0.85 + ], + "group_size": { + "5": 32, + "6": 32, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 1131826432, + "up_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9971376106535134, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "6": 128, + "8": 32 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1288992368, + "up_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9973211410014253, + "down_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.15, + 0.85 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1332028224, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9979512503272608, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1451303744, + "up_proj": { + "bits": [ + 8, + 6 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "6": 128, + "8": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9992382948924052, + "down_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "gate_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 1706107712, + "up_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.layers.9.self_attn": [ + { + "accuracy": 0.9776542751412642, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 133692416, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.05, + 0.95 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9791533601911444, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 137624576, + "v_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.25, + 0.75 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9811595929296393, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 141915264, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9867197745724728, + "k_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 3, + 2 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "2": 64, + "3": 64 + }, + "scale_bits": 4 + }, + "total_bits": 167308544, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9886607731643476, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197778944, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9889511944432008, + "k_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 128, + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 197944704, + "v_proj": { + "bits": [ + 4, + 3 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "3": 64, + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9932914744866522, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254402048, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9936180483353766, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 128 + }, + "scale_bits": 4 + }, + "total_bits": 254567808, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9941708751414952, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 64 + }, + "scale_bits": 4 + }, + "total_bits": 256707328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9944715727316705, + "k_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + }, + "total_bits": 260323328, + "v_proj": { + "bits": [ + 4 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "4": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.994361021801045, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 128, + "5": 128 + }, + "scale_bits": 4 + }, + "total_bits": 260859264, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9948085781775022, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 262998784, + "v_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9952251405308121, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 64, + "5": 64 + }, + "scale_bits": 4 + }, + "total_bits": 267385856, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 64 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9956507357327562, + "k_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 5, + 4 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "4": 32, + "5": 32 + }, + "scale_bits": 4 + }, + "total_bits": 271333376, + "v_proj": { + "bits": [ + 5 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "5": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.99735133957706, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 128, + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 328326656, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9978129512777454, + "k_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6, + 5 + ], + "bits_prop": [ + 0.1, + 0.9 + ], + "group_size": { + "5": 32, + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 334247936, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9982372420398813, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + }, + "total_bits": 380231168, + "v_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 128 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9988644308361568, + "k_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 6 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "6": 32 + }, + "scale_bits": 4 + }, + "total_bits": 396638208, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 32 + }, + "scale_bits": 4 + } + }, + { + "accuracy": 0.9995245128673943, + "k_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "o_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "q_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + }, + "total_bits": 506060288, + "v_proj": { + "bits": [ + 8 + ], + "bits_prop": [ + 1 + ], + "group_size": { + "8": 128 + }, + "scale_bits": 4 + } + } + ], + "model.norm.norm": null + } +} \ No newline at end of file