{ "last_module_idx": 58, "measurement": { "lm_head.linear": null, "model.layers.0.mlp": [ { "accuracy": 0.8554090073234156, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.860267162322998, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.8846290237025211, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.8924376462635241, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9291082557878996, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.934638079844023, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9454463783063387, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9621531900606657, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9658634725369906, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.963799304083774, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9683069618124711, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9815000577976829, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9841630443146354, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9897390870671523, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9905437991807335, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9928342652948279, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9970063437756739, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.0.self_attn": [ { "accuracy": 0.8732743388728091, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.8831450562728078, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9056574294441625, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9327854921943263, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9379436091372841, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9390872465936762, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.959879718328777, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9607762443391901, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9653665235168055, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9667352187006097, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9692553720976177, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9718346219313773, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9747779683062905, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9769451680936312, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9862089910005268, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9885595346751966, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9893585459182137, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9943474135116527, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9972075651350775, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.1.mlp": [ { "accuracy": 0.9216028451919556, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9257007335361681, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9425993153923437, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9505578969654284, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9602374277616802, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9634831579107987, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9726493891916776, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9781049882110796, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9803555152918163, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9795771492154974, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9818308604390997, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9892569192146, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9908489600608223, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9938012357605132, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9944357640649143, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9963175430893898, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9976449748011011, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.1.self_attn": [ { "accuracy": 0.8726400701623214, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.882448535216482, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.8951821389951204, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9345391486820421, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9363319434617695, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9372455948277524, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9592989243959126, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9605129235669186, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9637275369543779, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9654352288497121, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9680695878831964, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9689208080894068, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9719984186323065, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9731278168527704, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9847786724567413, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9871660392535361, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9893377713466945, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9934343997585146, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9971869056554217, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.10.mlp": [ { "accuracy": 0.9838263361077559, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9842866345455772, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9866297166598471, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9872535388720663, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9918373823165894, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9925468399336463, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9934771225640648, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9958335663142958, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9961668535282737, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9958143642074183, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9963598141544744, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9978456085449771, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9981674642155045, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988233594125823, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9988735270147261, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990595733060649, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996359437998188, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.10.self_attn": [ { "accuracy": 0.9782514023153406, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9789263010025024, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9843126017796365, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9887618171541315, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9890681881653635, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9892814739754325, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9941023929338706, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9943966689078432, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.994830178979196, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9949852079153061, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9945959766444407, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9949897472795687, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9961257773010355, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9964366186606256, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9978666952566096, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9981972216383407, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984563715559872, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991508632114059, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9995487991132235, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.11.mlp": [ { "accuracy": 0.9840682337158605, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.984466552734375, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9866543004387304, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9872648472848692, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9918377524928043, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.992595584768998, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9934967795484945, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.995781874970386, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9961650363708797, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9957824953292546, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.99635932006334, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9978225107647871, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9981614194418255, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988042626922068, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9988620399840569, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990448054710501, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996203502493077, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.11.self_attn": [ { "accuracy": 0.9758527514181639, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9766543068383869, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.982295141408318, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9875576786304775, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9877302662322396, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9880453536384984, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9933561055283797, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9936729179401147, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9941690909235101, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9943439007589692, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9939325569491637, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9943596834414884, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9955852866957062, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9959379830642751, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9975649196850626, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9979502846928019, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9982596929920348, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990345659224611, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.999494492664541, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.12.mlp": [ { "accuracy": 0.985206525576742, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9856312306303727, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9876694961598045, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9882240365994605, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9925068248259393, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9931991664986861, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9940148348871031, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9961675597648871, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964858744489519, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996147919250162, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9966642040955392, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9980117317877317, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9983203936564294, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9989100457414201, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.998962100105066, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.999122806383591, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996544503067669, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.12.self_attn": [ { "accuracy": 0.9791076057835629, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9798314508638883, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9847597558247415, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9888082209386324, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.989592067505184, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9898422893724943, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9943667312985972, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9946564882993698, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9950673321360036, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9952183648159629, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9948592154603255, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9952515112726312, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9963347743216314, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9966336116194725, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9979876707258978, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9983081245108655, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9985231956172931, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992051408381054, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9995661154879552, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.13.mlp": [ { "accuracy": 0.9844537455784647, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9848808865798148, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9871033516369367, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.987714304735786, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9921629570032421, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9928260100515265, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9937260264628812, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9959943561177504, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9963001577477706, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9959941847543967, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9965047989236681, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9979400287725424, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9982373553671335, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988586683022348, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.998922098329977, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990997502678319, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996360483285236, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.13.self_attn": [ { "accuracy": 0.9802039124463734, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9809562689379642, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9858320653438568, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9897649899909371, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.990005437480776, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9901823213225917, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9944541305303574, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9947144016623497, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9950967227157793, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9952383178629374, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9949465621458856, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9952919455735307, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964380946598554, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9967270142545825, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9980510163464045, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9983452431073314, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9985723489601361, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992353368158403, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9995808271985305, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.14.mlp": [ { "accuracy": 0.9861217846995906, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.986484728361431, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9884817372811469, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9890269911602924, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9929979459235543, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9935910709594425, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9943898355490283, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964194682083631, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9966963874666315, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964101718444573, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.996870977117827, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9981526567747718, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9984216404784667, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9989852454317244, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990315078512618, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.999191780133467, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996806796346056, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.14.self_attn": [ { "accuracy": 0.98098161346034, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9817721090818706, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9858798949342025, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9899886407350239, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9904671554502688, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9906360232516339, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9947831226806891, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.995015419235355, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9954104458815173, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9955633859101095, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9952612233005071, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9955893017743763, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964643014889014, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9967613018264896, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9980518233619238, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9983614244939465, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9986238709013713, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992150627076626, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9995922455809226, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.15.mlp": [ { "accuracy": 0.9872353625924963, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9875640931882357, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9894735444533197, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9900033238687014, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9935353829672462, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9940814587630724, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9948616494473658, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9966833007178808, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9969564323362551, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9966835328622868, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9971085641729204, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9982889257371426, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985401178465078, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990592309714932, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991035536911926, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992561176615326, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996968634230526, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.15.self_attn": [ { "accuracy": 0.9821410132081885, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.982662917752015, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9876838608791954, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9907569846040324, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9910333697733126, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9911803242407347, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9952899896000561, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9955387476243471, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9958790462268027, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9959982712017862, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9957408830523491, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9960077471638981, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9969804496749451, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.997232443996166, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9983388256085547, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985938968234941, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.998771294755371, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9993422569118833, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9996286670824415, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.16.mlp": [ { "accuracy": 0.9884685411265022, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9887498656385824, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9904787760031851, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9909510965410032, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.994180865585804, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9946693409430353, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9953691061390074, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970438558804361, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9972797567514997, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970202742046431, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9974015708032408, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984604841784427, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9986863422550654, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.99915919139197, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991905373079997, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9993215291515777, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9997320010964024, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.16.self_attn": [ { "accuracy": 0.9845258075939981, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9853028247230932, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9888362053193545, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9917436932262621, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9921198126516844, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9923028898866553, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9956657678673142, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9958488247112224, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9962281772964879, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9963393681927731, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9961297966931996, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9964045388133902, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9972317422691145, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9974782196314711, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984805178093282, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9987208050136503, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988761156993476, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9994014262368804, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9996594370489842, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.17.mlp": [ { "accuracy": 0.9881981305385891, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9884664016334634, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9902432827573073, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9907504014278713, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9939880367172392, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9944807467492003, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.995238202575006, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9969039310358072, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9971787747191755, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9969150631835586, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9973071307728165, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984108289017489, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9986396463293779, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9991256142721364, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991640378966143, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9993050130280224, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9997117106410626, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.17.self_attn": [ { "accuracy": 0.9869201795050973, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9874607865747652, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9905121389188265, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9931037567163769, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9934404974705294, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9935549019198668, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9962249602142134, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.996443393983339, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9967153191958603, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967984541466361, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967242336194766, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9969754007301832, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9976116316883188, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9978112809752163, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9986928819041503, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9988972585844366, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990282512613033, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9994777405242387, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9996972923706237, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.18.mlp": [ { "accuracy": 0.9877800925781852, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9880587999757967, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9898698620106045, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9903832438744997, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.993777124505294, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9942936846300175, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9950675423208036, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967937440072235, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970839164758983, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9968121867430838, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9972171130541124, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9983614251801842, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985959464193959, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.99910298156503, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991423560208396, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.999288582831229, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9997144971091888, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.18.self_attn": [ { "accuracy": 0.985008324447431, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9858139423947585, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9888369684156618, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9921135494583532, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9924790898435994, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9926087746494695, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9957813926433262, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9959450928788436, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9962709071604829, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964030800681365, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9962961116903707, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9965332711213514, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9971948577777335, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9974227022183569, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984616248035117, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9986975199886059, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988996598281359, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9993713599090513, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.999660570542083, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.19.mlp": [ { "accuracy": 0.9878526133926291, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9881382004210824, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9899482538825587, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9904726516259345, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9937857773743177, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9943035950786189, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9950877263357765, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967640983430963, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970656509854292, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996811814214054, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9972156211733818, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9983500203019694, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985878219533908, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990827838253034, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991330998508554, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992821620483148, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996998243671107, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.19.self_attn": [ { "accuracy": 0.9869843776288786, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9875102827423498, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9901413243067893, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9930035652298677, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9933778706349825, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9935140982270241, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.996345417672082, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9964890558468668, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9967690818011761, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9968712171049494, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967402229575735, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9969617695008454, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9975346583676966, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9977538587623521, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.998649255127499, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9988549166408024, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990327582743607, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9994379515808663, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9996882672176549, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.2.mlp": [ { "accuracy": 0.9624340816547996, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.963353621332269, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9692529439926147, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9709688550547549, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9807913585713035, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9824773998636949, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9850175600302847, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9901383801510459, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9910404525305095, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9901545432053114, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9914633655234387, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9949617499583646, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9957312051402895, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9972937079636675, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9974162266834786, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9979150534460419, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992554933322888, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.2.self_attn": [ { "accuracy": 0.9200100459550556, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.923262357711792, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9336897573973003, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9559247932936016, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.959876813386616, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9601389106951261, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9775135595547525, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9779836378599468, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9794622691054093, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9805722471914793, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9796506119401831, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9808492237015775, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9830513471051266, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9843915685227043, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9907420842271102, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9921398115785498, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9941053370896139, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9959789106720373, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984332226020726, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.20.mlp": [ { "accuracy": 0.9882152488357142, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9884369232152638, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9903753611602282, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9909959051169848, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9938794135263092, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9943609982728958, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9952631028074967, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996778507177767, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970972649753094, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9968454347629296, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9972365108367643, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9983610439261323, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985926017949456, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990843591329298, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991304939122576, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9993009629512304, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996904501397359, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.20.self_attn": [ { "accuracy": 0.988249797570078, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9885775548847098, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.991249038984901, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9932452855925811, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9940755731965366, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9942908937993803, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9967729239479491, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9969463767974001, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9972028142135394, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.997280121633881, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9971884073395478, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9973549062484189, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9979264104836866, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9980929246858546, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988736922019407, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990345288656259, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.999161295494751, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995354067926344, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997258766701347, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.21.mlp": [ { "accuracy": 0.9889174221377623, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9891266509106285, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9906930115662123, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9911631871210901, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9942989286623503, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9947570495699581, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9954598475443689, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970272597121564, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9973105374527605, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970727960922217, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9974313384216082, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984861349588946, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9986962277050081, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.99916264159899, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992030104622245, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9993346817791462, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.999731994209517, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.21.self_attn": [ { "accuracy": 0.988541061940946, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9890306090053759, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9914540272010001, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9932375714967125, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.994077021746259, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.994299557648207, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9967012252462538, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9969489766578925, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9972210487252787, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.997292259610013, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.997191892250588, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9973693511595851, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.99798386936125, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9981535706473025, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988994737596888, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.999066248438076, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9991530647599384, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995548717775627, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997273912212175, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.22.mlp": [ { "accuracy": 0.9878519255863992, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9880817932517905, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9898509438100614, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9904214065325888, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9936818229524713, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.994200815103556, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9950267194133056, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9966632182660856, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970124324685649, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967277383333758, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9971533607887594, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.998300780003008, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.998545247081079, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990533338369507, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990963957419521, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992497750116807, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996840621001626, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.22.self_attn": [ { "accuracy": 0.9873030875858507, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9878112485534266, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9908757688183534, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9927451179215783, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9936081398474542, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9937707166138449, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9966264704340383, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9968118114691031, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9971022449041668, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9971513859927654, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.99703128479029, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9972587494473708, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9978393561353809, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9980526478274873, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988151574017186, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990106918113796, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9991102258821851, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995068621968752, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997098655930083, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.23.mlp": [ { "accuracy": 0.9877955646891343, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9880411444525969, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9897431260661075, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9902623782032415, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9937424596987272, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9942307264397019, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9949931149419985, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967425006784891, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.997043079059375, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9967893134606512, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9971811042020196, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9983429351919576, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985721137766775, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990866205802089, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991294958285595, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992767626321629, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9997129158950165, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.23.self_attn": [ { "accuracy": 0.989171575558813, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9892956810562235, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9917997380620555, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9934561836876368, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9944169313499802, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9946524347913893, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9970622466583001, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9972042108052656, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9974737822225219, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9975831616473826, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9974020621494243, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9976196149854284, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9981069719712985, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9982706925978786, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9989566167718485, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991296186651054, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9992440022332104, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995594904022781, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997507422231138, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.24.mlp": [ { "accuracy": 0.9869069323727959, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9871694908330315, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9890265009905163, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9896003635306108, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9932861610462791, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9938001036643982, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9946385947497267, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9965035264429293, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996827154959503, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9965588768061838, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9969763800893959, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9982277907823262, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9984700369011414, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9990247770359642, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990719028405453, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992355050421075, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996981608779415, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.24.self_attn": [ { "accuracy": 0.9877640118724421, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9882522382234272, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9911435034714247, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9931276456305855, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9940445379991281, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9942569489541807, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9967375887851966, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9969107851778206, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9971837605300703, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9972968958318233, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9971426750876402, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9973561595145025, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9978466688802368, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9980438349670485, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988110556610321, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9989985260916384, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9991388797367874, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995020560332035, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997154091190743, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.25.mlp": [ { "accuracy": 0.9844199340594443, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9847083970120079, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9869826898763054, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.987701490521431, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9919718456895728, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9926057355968576, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9936346047018704, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9957940249066604, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9961932744634779, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.995879452479513, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9963859456934427, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9978726970914163, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9981729650968, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988251183378068, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9988881680918368, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990922718455917, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996323481594261, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.25.self_attn": [ { "accuracy": 0.9860665672703793, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9874674128858667, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9902763429440951, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9920278513117841, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9936031107055513, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9935549293693743, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9968098158899107, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9969829009159615, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9972427524626255, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9973239071275058, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970418830451212, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9973757319936627, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9979150936399636, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9980442137702515, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.998866120549409, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990134674467539, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9991835126359212, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995046695694327, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997244111605381, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.26.mlp": [ { "accuracy": 0.9821718397893404, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9825612526190909, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9853630254143163, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9862472211059771, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9908847118678846, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9915879196242282, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9928483884585532, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9952305535736837, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.995679826328629, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9953250783054453, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9958947720496278, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9975914184592272, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9979290742623178, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9986652803460234, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9987476290644783, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9989982994371339, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995857055151933, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.26.self_attn": [ { "accuracy": 0.9865743408077642, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.987444782727643, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9896076266702852, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9917201297847849, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9930510983655327, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9934074376758776, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.996964375439443, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.997071956725497, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9972954170876428, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9974129172532182, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9972778086207414, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9974935117520785, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9977997535545575, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9979557247930452, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9987809172782459, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9989824899913449, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9991731024102161, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9994632782610623, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997192056310412, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.27.mlp": [ { "accuracy": 0.9707535599407396, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9715930505802757, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9764203868414226, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9777438954303139, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9859795601744401, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9870938760669608, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9889673763199857, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9927537841232199, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9933369802801233, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9927777436218763, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9936631047411969, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996272894896959, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9967905112395161, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9979386959028872, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9980347156524658, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9984121065782873, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9993391650679865, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.27.self_attn": [ { "accuracy": 0.9879381288039056, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.989866557089906, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9919138503702063, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9941020623633736, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9949063466567742, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9950317269877383, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9971556706648124, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9972514184681993, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9975295209963071, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9976573686458563, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9975666684147558, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9977913877289546, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.998002590709611, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9981965959856385, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9989006735972668, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.999062740979226, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9992250461425436, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9994935890855757, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9997274848436447, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.3.mlp": [ { "accuracy": 0.9647314109300312, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9654904854925055, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9708613784689653, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9724296896081221, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9819360673427582, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.98346545821742, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9858002348950035, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9906850323865288, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9915501436120585, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9907386075509222, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9919297804957942, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9952584387440431, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9959673430574568, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9974581630606401, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9975743229059797, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9980468889208216, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9992962546254459, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.3.self_attn": [ { "accuracy": 0.9299100072760331, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9324928898560374, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9461031650242052, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9638102493788067, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9649033734672948, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9653886870334023, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9811589921775618, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9817438800083963, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9829602806191695, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9835786427322187, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9823948427250511, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9835263522047746, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9865173854325947, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9875228279515317, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9926980608387997, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9937506180844808, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9950935691595078, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9969925196155122, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9986976868423977, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.4.mlp": [ { "accuracy": 0.9717588518795214, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9724319859554893, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9767575765910902, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9779856361840901, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9855788224621823, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9868412096249429, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9886777181374399, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9925753435021952, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9932565226366645, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9925880283117294, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9935706834259787, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9961991706176808, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9967813262421834, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9979543125158862, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9980439822140493, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.998412680861197, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9994216418678039, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.4.self_attn": [ { "accuracy": 0.9549308795677989, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9565280707258927, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9645941979006717, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9770096289484125, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9774391776637027, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9776882748854788, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9878371028523696, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9882113737495322, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9889965739689375, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9894893679179644, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9886341165555151, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9893706088003359, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9910798378680882, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9917857788111034, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9951152777985522, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9958569572160119, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996823610443818, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9979958089166566, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9991342930221244, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.5.mlp": [ { "accuracy": 0.9761800954216405, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9768167357695731, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9804606123974449, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9814633821186266, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9879509696834966, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9889864121612749, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9904760326209822, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9938317894151336, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9943413726593319, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9938169647204248, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9946204937602344, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9968269898703224, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9973045618910539, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9982793958563554, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9983577306725477, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9986557039971414, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995012597033852, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.5.self_attn": [ { "accuracy": 0.9604767655071459, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9618853173757854, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9697414511128476, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.979678246535753, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9801261001511624, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9804188405212603, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9893247814554917, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9896571416603891, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9903339155410465, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9906899223202154, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9900463090131157, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9906631504234514, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9923930066196542, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9929709575678173, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9958557831613641, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9964699674593774, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9972180051631049, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9983137863639154, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9992358567389218, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.6.mlp": [ { "accuracy": 0.9780441255945909, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9786776978718607, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9820381199058733, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9829397844640833, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9889469554549769, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9899034264840578, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9912542238047248, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9943627381795331, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.994822632325323, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9943292050769454, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9950728498791394, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9970917658586251, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9975313422711272, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984261655297718, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9984922241419554, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9987614234223178, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9995327998246801, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.6.self_attn": [ { "accuracy": 0.9642330847288433, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9655292943904275, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9728665163642481, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9817885913346943, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9821537968359495, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9824489637425071, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9904710689657613, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9908023633454975, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9914140379742572, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9917106024528805, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9911306806300816, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9916797672447405, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9932602094976526, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.993773735275394, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9963311423596583, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.996870401266374, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9975131291307902, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985053214783731, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9993084279428187, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.7.mlp": [ { "accuracy": 0.978815064618462, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9793789480861864, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9826750127892745, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9835932944950304, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9892606076441313, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.990211609947054, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9915625421624434, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9944803353987242, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9949473526917005, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9944692650123647, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9952141097501704, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9971537497874937, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9975861291352072, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9984437330184799, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985241726236908, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9988093441842418, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.99952488108293, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.7.self_attn": [ { "accuracy": 0.9656342989520023, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9668074030625193, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9737494682010851, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9823550189796247, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9827180802822113, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9830116535487928, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9907590958632921, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9910880162527687, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.991688913420627, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9919953730545545, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9913974937639738, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9919481348050269, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9934616449632143, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9939566446762336, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964307330940899, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9969644472002983, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9975927481917959, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9985489321774558, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9993281768340814, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.8.mlp": [ { "accuracy": 0.9848931086690802, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9853568202570865, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.987495503927532, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9880542315934834, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.992434068730003, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9930736751932847, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9939052897848581, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996140214957689, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9964268654584885, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996119558419052, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9966190414209115, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9980017221287677, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9983011868439222, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9989041978199231, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9989555372219336, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9991241219012361, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996543751637402, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.8.self_attn": [ { "accuracy": 0.9696789697596901, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9708416022752461, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9770174873502631, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9844938140166434, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.984914851816077, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9852150929601569, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.99180836740293, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9920773764974192, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9926198093514693, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9928823929083975, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9925150910490438, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9929688910120412, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9942892636907729, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9947079027953901, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.996876803079718, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9973210098319932, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9978376338748556, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9986678595213514, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9993819608107993, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.9.mlp": [ { "accuracy": 0.9841843184671903, "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": 169638736, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9846550414436742, "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": 175405904, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9869073001961959, "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": 195969536, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9875095039606094, "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": 219857408, "up_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.3, 0.7 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9920997258863951, "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": 248018656, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9927414914495066, "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": 254268416, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9936303594394734, "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": 273355232, "up_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "3": 32, "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9959720898615686, "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": 312411232, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9962626648576636, "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": 317019136, "up_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9959490118842376, "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": 322238176, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 128, "5": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9964521903740732, "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": 328487936, "up_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9979128359179747, "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": 396457696, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 128, "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9982148246153405, "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": 402707456, "up_proj": { "bits": [ 6, 5 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "5": 32, "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9988596836399091, "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": 458487520, "up_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9989094419502899, "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": 476255488, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9990912931726167, "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": 518722816, "up_proj": { "bits": [ 8, 6 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "6": 128, "8": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9996495024350128, "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": 606803200, "up_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.layers.9.self_attn": [ { "accuracy": 0.9751499859910262, "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": 53879808, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.05, 0.95 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9759666009953147, "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": 55387136, "v_proj": { "bits": [ 3, 2 ], "bits_prop": [ 0.25, 0.75 ], "group_size": { "2": 64, "3": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9815236188863453, "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": 58040192, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9871609508991241, "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": 67377920, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9873219116737968, "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": 79827456, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 128, "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9875857673193279, "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": 79861376, "v_proj": { "bits": [ 4, 3 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "3": 64, "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9930991232395172, "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": 101847552, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9933738151663228, "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": 101947008, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9939036071300507, "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": 102837504, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9941141574006331, "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": 104220672, "v_proj": { "bits": [ 4 ], "bits_prop": [ 1 ], "group_size": { "4": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9936482878892046, "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": 105027200, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 64, "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9941222522603838, "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": 105458944, "v_proj": { "bits": [ 5, 4 ], "bits_prop": [ 0.1, 0.9 ], "group_size": { "4": 32, "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9953485390073374, "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": 108078080, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 64 }, "scale_bits": 4 } }, { "accuracy": 0.9957244486400956, "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": 109660160, "v_proj": { "bits": [ 5 ], "bits_prop": [ 1 ], "group_size": { "5": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9974393924992335, "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": 132911616, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9978441950914106, "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": 134825984, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9981952695862243, "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": 152179200, "v_proj": { "bits": [ 6 ], "bits_prop": [ 1 ], "group_size": { "6": 128 }, "scale_bits": 4 } }, { "accuracy": 0.9989665678065074, "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": 160843776, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 32 }, "scale_bits": 4 } }, { "accuracy": 0.9994801930887135, "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": 202510848, "v_proj": { "bits": [ 8 ], "bits_prop": [ 1 ], "group_size": { "8": 128 }, "scale_bits": 4 } } ], "model.norm.norm": null } }