{
"in_dir": "base_model",
"out_dir": "quant",
"cal_dataset": "wikitext-test.parquet",
"bits": 4.0,
"dataset_rows": 100,
"measurement_rows": 16,
"length": 2048,
"measurement_length": 2048,
"head_bits": 6,
"shard_size": 8192,
"compile_full": null,
"rope_scale": null,
"rope_alpha": null,
"output_measurement": null,
"progress": "quant",
"status_output": false,
"cal_filename": "quant/cal_data.safetensors",
"last_module_idx": 66,
"measurement": {
"model.layers.0.self_attn": [
{
"accuracy": 0.8954753950238228,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.915768027305603,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9205832555890083,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9524500295519829,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9533471055328846,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9538668505847454,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.966143298894167,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9675611928105354,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9706969987601042,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9724404159933329,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9762455597519875,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775304272770882,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9787847846746445,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9804443344473839,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9882887806743383,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9900966100394726,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9907924979925156,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.994104434736073,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975414818618447,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.0.mlp": [
{
"accuracy": 0.859587624669075,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.8620419800281525,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.8814472854137421,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.8879095986485481,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9379775263369083,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9440193064510822,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.953786414116621,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9683481156826019,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9714913945645094,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9692010469734669,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9732353985309601,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9843819513916969,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9867362882941961,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9914362244307995,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920876501128078,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9942444544285536,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975855492521077,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.1.self_attn": [
{
"accuracy": 0.8828793317079544,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.889796756207943,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9061766564846039,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9298331290483475,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9319383352994919,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.942933700978756,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9411077499389648,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9548083506524563,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9665140472352505,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9682299047708511,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9702100902795792,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9756643157452345,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9722584132105112,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9781700279563665,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9829313959926367,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9886683244258165,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9851147020235658,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943198119290173,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962473693303764,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.1.mlp": [
{
"accuracy": 0.9055621400475502,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9164061024785042,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9247237294912338,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9251643791794777,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9810757022351027,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850738467648625,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871309101581573,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9914046805351973,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924698853865266,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924200172536075,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930565226823092,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.996183990733698,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964683833532035,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978015893138945,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978882826399058,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983617896214128,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988936479203403,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.2.self_attn": [
{
"accuracy": 0.9662072509527206,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9703861568123102,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9798554629087448,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.983360517770052,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9849622352048755,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9859323026612401,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9888906618580222,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.989962044171989,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906571777537465,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.990931786596775,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925067881122231,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934602831490338,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9938895469531417,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947602623142302,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966423606965691,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972789641469717,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970852490514517,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988314043730497,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992141086841002,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.2.mlp": [
{
"accuracy": 0.9670207314193249,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9678649716079235,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9725988544523716,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9740983378142118,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9834451042115688,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.98478330578655,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9868618585169315,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9913856172934175,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9922013450413942,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915548954159021,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9926309231668711,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956863140687346,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963247945997864,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976629447191954,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978000235278159,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982310999184847,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.999374316597823,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.3.self_attn": [
{
"accuracy": 0.9574964717030525,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9608141295611858,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9692790992558002,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9749292340129614,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9784314911812544,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9787397272884846,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833304826170206,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9851747648790479,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9872469566762447,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9875043900683522,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9897161722183228,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910929119214416,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.991242635063827,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927708064205945,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.995061969384551,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963864993769675,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.995709721930325,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983382426435128,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988283232087269,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.3.mlp": [
{
"accuracy": 0.9540744684636593,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9551870301365852,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9620404839515686,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9641680866479874,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9767712149769068,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9786598309874535,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9816190209239721,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9879348734393716,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9890726748853922,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9881607508286834,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9896537084132433,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.993955900426954,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948404673486948,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967302805744112,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969191020354629,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975319148506969,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991283424315043,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.4.self_attn": [
{
"accuracy": 0.9554185308516026,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9607311263680458,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.968228243291378,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9740068707615137,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9761759415268898,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9765117075294256,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9814548920840025,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9835245814174414,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9862462906166911,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9868802968412638,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9883255548775196,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9900972452014685,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9905680622905493,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920551627874374,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.994685921818018,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960050634108484,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953980729915202,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.998075638548471,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987751066219062,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.4.mlp": [
{
"accuracy": 0.9416235722601414,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9431620351970196,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9521548114717007,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9548864178359509,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9704616293311119,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9729410167783499,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9767039008438587,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9846104038879275,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9860550938174129,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9849321562796831,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9868718720972538,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923016233369708,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934483012184501,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958080495707691,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960745670832694,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968755342997611,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988737220410258,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.5.self_attn": [
{
"accuracy": 0.9440676532685757,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9479215443134308,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.956732552498579,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9646333903074265,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9706690926104784,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9730441439896822,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9788974653929472,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.980994576588273,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9835685044527054,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9851872390136123,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9861069703474641,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.987149897031486,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9886868894100189,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9899836145341396,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933451521210372,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948798594996333,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9944253922440112,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974782518111169,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985776115208864,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.5.mlp": [
{
"accuracy": 0.9273791685700417,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9293689206242561,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9405047297477722,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9437909796833992,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9634025655686855,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9664154648780823,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9710105657577515,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9809325225651264,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9826938677579165,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9813214149326086,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9837073050439358,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9904730431735516,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918760657310486,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948123972862959,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951465418562293,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.996118419803679,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985968837281689,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.6.self_attn": [
{
"accuracy": 0.9413569122552872,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9444054216146469,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9528472796082497,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9622189626097679,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9692628420889378,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9708127472549677,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9785673934966326,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9793477300554514,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9824373573064804,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833459351211786,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9849908472970128,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9864804223179817,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9869887894019485,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894773187115788,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9926262060180306,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9945895094424486,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937369031831622,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972283514216542,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984910697676241,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.6.mlp": [
{
"accuracy": 0.919913150370121,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9219547659158707,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9350977763533592,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9390667043626308,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9594808742403984,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9628869369626045,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9683231562376022,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9788286294788122,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9808528944849968,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9793181903660297,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9819846339523792,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894404327496886,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910067543387413,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9942439212463796,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946284047327936,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958166372962296,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984571194509044,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.7.self_attn": [
{
"accuracy": 0.933421678841114,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9355790689587593,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9441087208688259,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.956241812556982,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9653680622577667,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9668463468551636,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9752421863377094,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9767936300486326,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9801823571324348,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9822937529534101,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9838634617626667,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9848697707056999,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9860667428001761,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9874788299202919,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921919200569391,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937374326400459,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940313128754497,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963794697541744,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985294949728996,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.7.mlp": [
{
"accuracy": 0.9132847636938095,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9156109988689423,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.929191492497921,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9332743361592293,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9562145099043846,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.959967065602541,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9654836431145668,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9772119633853436,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9793235417455435,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.977599723264575,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9805434942245483,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9885309943929315,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.990282392129302,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937539156526327,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941439051181078,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953300678171217,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983309059171006,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.8.self_attn": [
{
"accuracy": 0.9345854669809341,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9383742660284042,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9450973570346832,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9576957561075687,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9651788398623466,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9680792614817619,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9750993922352791,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9786972030997276,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9806760288774967,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9815903194248676,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9828728064894676,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9847475737333298,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.984885380603373,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9874300016090274,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9914852948859334,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933625590056181,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9932516235858202,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962233870755881,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981715388130397,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.8.mlp": [
{
"accuracy": 0.9094923660159111,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9118784219026566,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9262440651655197,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9305976182222366,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9542099311947823,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9580841027200222,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9640124142169952,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9761284701526165,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9783470071852207,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9765878617763519,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9796071890741587,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880139091983438,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.989814325235784,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934720220044255,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9938792502507567,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951762263663113,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982445405330509,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.9.self_attn": [
{
"accuracy": 0.93181112408638,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9341230094432831,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9416931606829166,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9545429423451424,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.964818187057972,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9657629169523716,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9762662090361118,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775382168591022,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9790038000792265,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9810895267874002,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9824847746640444,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9839906841516495,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850038532167673,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9866630798205733,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9913717247545719,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933139220811427,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935893495567143,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.996285465778783,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984261628706008,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.9.mlp": [
{
"accuracy": 0.9065756276249886,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9090653285384178,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9241945594549179,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9288148209452629,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9525704905390739,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9566186852753162,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9629656486213207,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9752628058195114,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775536004453897,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9757172130048275,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9788635708391666,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9875574233010411,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894396215677261,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9932326218113303,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936450002714992,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950094623491168,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981947726337239,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.10.self_attn": [
{
"accuracy": 0.9245620742440224,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9273806065320969,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9339804202318192,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9482939541339874,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9613415636122227,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9629591405391693,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9738142862915993,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.975861107930541,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9780932702124119,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.978845402598381,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9812089614570141,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.982582924887538,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9835674557834864,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9848895249888301,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9907076256349683,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921699417755008,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930778727866709,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953026999719441,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981228365795687,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.10.mlp": [
{
"accuracy": 0.9055445790290833,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9081645235419273,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9242616444826126,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9292544648051262,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9519640058279037,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9560526423156261,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9628417566418648,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9746240321546793,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9771080315113068,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9753909297287464,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.978566812351346,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9874020088464022,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9892825186252594,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930769158527255,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935754807665944,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950704132206738,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981404969003052,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.11.self_attn": [
{
"accuracy": 0.9210747480392456,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9247877299785614,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9323282316327095,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9480321742594242,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9578948989510536,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9595053195953369,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9705150723457336,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.972735183313489,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9764079917222261,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.97774906642735,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9797142539173365,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.980163810774684,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.98124098777771,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9836530368775129,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9898796211928129,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.991933386772871,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923942694440484,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952917327173054,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979786253534257,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.11.mlp": [
{
"accuracy": 0.9062269255518913,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9088773727416992,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9248028993606567,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9297933131456375,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9521685354411602,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.956260722130537,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9630573131144047,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9746693521738052,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9771793968975544,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9755045436322689,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.978668512776494,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.987466536462307,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.989334050565958,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930931534618139,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936163630336523,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951287405565381,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.99815122585278,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.12.self_attn": [
{
"accuracy": 0.9176041781902313,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9214098230004311,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9289238080382347,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.943946361541748,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9578831680119038,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9590032137930393,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9713136237114668,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9726635236293077,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9751110654324293,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.976806528866291,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9786937013268471,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9803492054343224,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9812966100871563,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.983193214982748,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9892378710210323,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9914765367284417,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919486846774817,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952149474993348,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980315575376153,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.12.mlp": [
{
"accuracy": 0.9043113961815834,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9071694985032082,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9228239431977272,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9277072250843048,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9512331001460552,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.955496110022068,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9621148854494095,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9742928948253393,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9767845403403044,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9749837387353182,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9782731682062149,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871802199631929,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9891378665342927,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9929616595618427,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934605280868709,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949045581743121,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981096144765615,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.13.self_attn": [
{
"accuracy": 0.9234198108315468,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9266788512468338,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9334149062633514,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9502020925283432,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.960641585290432,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9619468227028847,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9747657738626003,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9763126689940691,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9786004107445478,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9795366823673248,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9805781673640013,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9816244058310986,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9828913286328316,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9843263290822506,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9903721353039145,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.991931576281786,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9932008967734873,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952920451760292,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982003853656352,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.13.mlp": [
{
"accuracy": 0.8991779312491417,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.902088537812233,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.918775886297226,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9241109415888786,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.948443129658699,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9529099650681019,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.960031308233738,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9726768247783184,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9753897320479155,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9735988657921553,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9770406149327755,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9864852000027895,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9885214120149612,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925489579327404,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930998315103352,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946833308786154,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979912317357957,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.14.self_attn": [
{
"accuracy": 0.9154813885688782,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9200187176465988,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9285646229982376,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9456786401569843,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9566418863832951,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9581002183258533,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9717060700058937,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9739729519933462,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9757568389177322,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9769398123025894,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9779077619314194,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.979666393250227,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9810224734246731,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9831309802830219,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9896517917513847,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915214339271188,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924991424195468,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952121274545789,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979641155805439,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.14.mlp": [
{
"accuracy": 0.8986156731843948,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9015168622136116,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9181070551276207,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9233970791101456,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9482467360794544,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9527352079749107,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9597132131457329,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9724361449480057,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9751418121159077,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9734549429267645,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9768946543335915,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9863980934023857,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884347375482321,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924465254880488,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930713637731969,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946370986290276,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979454386048019,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.15.self_attn": [
{
"accuracy": 0.9125399440526962,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9156104326248169,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9255928918719292,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9402788653969765,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9539141245186329,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9552047699689865,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9670593440532684,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9697687923908234,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9748417437076569,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9758072830736637,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775287080556154,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9796125739812851,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9805984869599342,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9828801974654198,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9887404153123498,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9911710321903229,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912800388410687,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948631743900478,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978609376121312,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.15.mlp": [
{
"accuracy": 0.90084557980299,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.903657466173172,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9192847535014153,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9243183508515358,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9495509900152683,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9537828452885151,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9604237638413906,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9732636008411646,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9758126344531775,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9742035027593374,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775092992931604,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9868235196918249,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9887647274881601,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927182686515152,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9932924113236368,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947318756021559,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.998009311966598,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.16.self_attn": [
{
"accuracy": 0.9225006625056267,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9260915070772171,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.93379345536232,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9479106366634369,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.959874652326107,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9608021304011345,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9728347025811672,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9743041526526213,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9773157741874456,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9797396399080753,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9806866683065891,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9821058828383684,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833320062607527,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850300783291459,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9905426744371653,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9922630563378334,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930246588774025,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954897677525878,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982016739668325,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.16.mlp": [
{
"accuracy": 0.9025580137968063,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9053507596254349,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9206036999821663,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9256125837564468,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9504885338246822,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9546359032392502,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9611783102154732,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9736590534448624,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.976141382008791,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9747020099312067,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9779549296945333,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9870770135894418,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9889909261837602,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9928160505369306,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934353209100664,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948184993118048,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980210568755865,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.17.self_attn": [
{
"accuracy": 0.9257747307419777,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9286039620637894,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.935335710644722,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9512667618691921,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9604676812887192,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9628506451845169,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9733078256249428,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9761419128626585,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9785907063633204,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9785222969949245,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9806365557014942,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9821312110871077,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9829443246126175,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9845194080844522,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9900851286947727,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.992334523703903,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925823672674596,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957830985076725,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.998045910615474,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.17.mlp": [
{
"accuracy": 0.8998426496982574,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9027659371495247,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9188177287578583,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9243087396025658,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9488994739949703,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9532327614724636,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9602273255586624,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9726381897926331,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9753408562391996,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9738565403968096,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.977249514311552,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9866273971274495,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9886263851076365,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925329461693764,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.993216899689287,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947362639941275,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979504551738501,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.18.self_attn": [
{
"accuracy": 0.930207721889019,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9331731498241425,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9395735412836075,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9516260400414467,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.962639968842268,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9652673155069351,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9730946496129036,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9768130294978619,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9795126505196095,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.980476139113307,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9825235716998577,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.983674393966794,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9845744278281927,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9859072640538216,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9899446750059724,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930429067462683,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916632119566202,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961132404860109,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979518209584057,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.18.mlp": [
{
"accuracy": 0.9042697697877884,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9071269109845161,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9222597181797028,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9276817291975021,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9511814154684544,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9553046114742756,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9620116390287876,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9736417233943939,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9762427378445864,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9750394895672798,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9782604668289423,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9872453026473522,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9891353575512767,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9928120351396501,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935439638793468,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949975581839681,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980023549869657,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.19.self_attn": [
{
"accuracy": 0.9376803860068321,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9405889175832272,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9459853880107403,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9577822908759117,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9668755829334259,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9686102047562599,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9771732855588198,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9793982692062855,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9815273508429527,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9830815084278584,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9844017662107944,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9854886159300804,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9862198662012815,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9876099163666368,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920846801251173,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937541531398892,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940847735852003,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965364616364241,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984150449745357,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.19.mlp": [
{
"accuracy": 0.9114648401737213,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9141155779361725,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9276958331465721,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9327016323804855,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.954801544547081,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9586106240749359,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9647428803145885,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9755066335201263,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9778970573097467,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9767964482307434,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9797745477408171,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880999755114317,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9898764984682202,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9932886897586286,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9939553877338767,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952727667987347,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981128118233755,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.20.self_attn": [
{
"accuracy": 0.9451662451028824,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9478352703154087,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9536472968757153,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9646199718117714,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9709294755011797,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.972251521423459,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.979300107806921,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9821022171527147,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9841908924281597,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9848207533359528,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9862387832254171,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871250810101628,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9878906579688191,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.989013091661036,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9928586450405419,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.994474679697305,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9944698577746749,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970161328092217,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985297875246033,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.20.mlp": [
{
"accuracy": 0.9196527525782585,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9219760745763779,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9336907863616943,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9378717355430126,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9592782184481621,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9625161550939083,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9677678197622299,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9782731980085373,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9802852533757687,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9792443942278624,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9818413183093071,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894269797950983,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9909407775849104,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941096259281039,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946414195001125,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.995749550871551,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983608438633382,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.21.self_attn": [
{
"accuracy": 0.9519269466400146,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9541482180356979,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9593300744891167,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9685815423727036,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9741687197238207,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9760780688375235,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9816852174699306,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9842663295567036,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9860157333314419,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9864440727978945,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9875335022807121,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9887787150219083,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.989285477437079,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9902572967112064,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937353101558983,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951940136961639,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950804025866091,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.997463630978018,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986839422490448,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.21.mlp": [
{
"accuracy": 0.9258313998579979,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9279342070221901,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9381502605974674,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9417422972619534,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9625239335000515,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9654535986483097,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9700090549886227,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9801121074706316,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9819174204021692,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9809171967208385,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9832717832177877,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.990284544415772,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916622722521424,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946128888987005,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950710856355727,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960036291740835,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984954033279791,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.22.self_attn": [
{
"accuracy": 0.959535475820303,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.961449034512043,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9675278924405575,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.972922021523118,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9778671283274889,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9799151513725519,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9831464663147926,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9856763742864132,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.988006173633039,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884339468553662,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9901307374238968,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9909731522202492,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912645034492016,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923930424265563,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946692027151585,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961016052402556,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954509604722261,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979514463339001,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988511857809499,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.22.mlp": [
{
"accuracy": 0.9281870052218437,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9301374405622482,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.93976990878582,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9430716559290886,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.963589932769537,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9664384536445141,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.970716642215848,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9807633552700281,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9825082197785378,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.981462299823761,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9837476052343845,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9905694015324116,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919005604460835,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947976809926331,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952179724350572,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960874207317829,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.998557236045599,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.23.self_attn": [
{
"accuracy": 0.9585595987737179,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9608400948345661,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9661816619336605,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9731915071606636,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775797836482525,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9791795592755079,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9834931325167418,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9861969761550426,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880285616964102,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884001696482301,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.989586197771132,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.990400604903698,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910352667793632,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919645925983787,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947002888657153,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960017874836922,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.995765304658562,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979711405467242,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988897367147729,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.23.mlp": [
{
"accuracy": 0.9299747124314308,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9318723306059837,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9409330151975155,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.944084458053112,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.964467242360115,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9672388434410095,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9712910130620003,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9812156781554222,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.98293093778193,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9819148257374763,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9841319732367992,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9908028859645128,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920955095440149,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949257811531425,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953309698030353,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.996146634221077,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985819715075195,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.24.self_attn": [
{
"accuracy": 0.9613390564918518,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9630597792565823,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9683404490351677,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9739083237946033,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9790964648127556,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9808864668011665,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9851742628961802,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871780555695295,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880079943686724,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9888165555894375,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9902446689084172,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910678016021848,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915370875969529,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.99261858407408,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950777771882713,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962649836670607,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.995957788079977,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978881615679711,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989249646896496,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.24.mlp": [
{
"accuracy": 0.9314836263656616,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9332762360572815,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9419373907148838,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9449307918548584,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9652171991765499,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9679170474410057,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.971783559769392,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9816436991095543,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9832949731498957,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9822856783866882,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9844612330198288,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.99097790569067,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9922561119310558,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950355920009315,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954197509214282,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961929724086076,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986155134392902,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.25.self_attn": [
{
"accuracy": 0.9599181339144707,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.961070392280817,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.966218389570713,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9723872523754835,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9781264942139387,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9797741249203682,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9837831445038319,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9861332289874554,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9881806410849094,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9886278240010142,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9899702398106456,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906275244429708,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912404688075185,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920991975814104,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948351122438908,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961124330293387,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957734639756382,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.997807146050036,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988794868113473,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.25.mlp": [
{
"accuracy": 0.9326912090182304,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9344697594642639,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9426952414214611,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9455682411789894,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9658484682440758,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9685135409235954,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9721763171255589,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9818990230560303,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9835778325796127,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9826094601303339,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9847552618011832,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9911602130159736,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.992411797400564,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951151604764163,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955186094157398,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962609075009823,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.998642495367676,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.26.self_attn": [
{
"accuracy": 0.9578017368912697,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9600430242717266,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9646002352237701,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9706894978880882,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9772729743272066,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.979037256911397,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.982847249135375,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850249933078885,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9872760577127337,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9881802806630731,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894789671525359,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906616946682334,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910595016553998,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918741276487708,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9945916477590799,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9959096955135465,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955893456935883,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975593676790595,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988381142029539,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.26.mlp": [
{
"accuracy": 0.9341814294457436,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.935901552438736,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.943786907941103,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9465231895446777,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9665994495153427,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9692160487174988,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9727182406932116,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.982348058372736,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9839572068303823,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.983004953712225,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850921994075179,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9913571141660213,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925679014995694,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952348242513835,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956188960932195,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963284400291741,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986768561648205,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.27.self_attn": [
{
"accuracy": 0.9652378559112549,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9664778001606464,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9718529507517815,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9779752101749182,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.981735548004508,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9827145505696535,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.987411635927856,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9883653931319714,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9897706126794219,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.990246762521565,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9911144440993667,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918317189440131,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924692469649017,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9931291411630809,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956633560359478,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965402795933187,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966327690053731,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981975591508672,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991107991081662,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.27.mlp": [
{
"accuracy": 0.9355442523956299,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9373545423150063,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9447178617119789,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9473666660487652,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9672805592417717,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9698762781918049,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9731602407991886,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9825671128928661,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9841927867382765,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833421912044287,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9853988084942102,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915208658203483,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927232363261282,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952938514761627,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956975691020489,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963598265312612,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.998678374919109,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.28.self_attn": [
{
"accuracy": 0.9586983695626259,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.960369236767292,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9650076329708099,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.973130414262414,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9786506332457066,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9792791493237019,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9851186694577336,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9862402472645044,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9879138302057981,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884906522929668,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9895032625645399,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9903269670903683,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9908324750140309,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9917955519631505,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947909428738058,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.995857804082334,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960249224677682,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976427555084229,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989375841105357,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.28.mlp": [
{
"accuracy": 0.9346529692411423,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9366000294685364,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9438350945711136,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.946636475622654,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9667219556868076,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9693902134895325,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.972698800265789,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9820746257901192,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9837354961782694,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9830381479114294,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9851469472050667,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9913496300578117,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925739848986268,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951255135238171,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956200839951634,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963067171629518,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986079147784039,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.29.self_attn": [
{
"accuracy": 0.953386127948761,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9553359150886536,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9615236110985279,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9702148027718067,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9758381862193346,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.976599371060729,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9839132353663445,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850270608440042,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9865177692845464,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9868638897314668,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9882249888032675,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.989140385761857,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.990016502328217,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9908444015309215,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943150063045323,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.995407959446311,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956911932677031,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.997375956736505,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988253758056089,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.29.mlp": [
{
"accuracy": 0.9342653304338455,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.936233215034008,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9436625018715858,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.946524266153574,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9664524868130684,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9691478870809078,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9726137388497591,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9819028787314892,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9835824817419052,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9828450009226799,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9849793901666999,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912399016320705,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.992491683922708,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950690097175539,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955552835017443,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962757660541683,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985875469865277,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.30.self_attn": [
{
"accuracy": 0.9528799466788769,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9554436132311821,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9611076563596725,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9691898208111525,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9762859866023064,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775823634117842,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833487756550312,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9848042409867048,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9864199170842767,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.987317749299109,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9885307056829333,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894011924043298,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9900705330073833,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9909136239439249,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943923144601285,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.995469419285655,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957389938645065,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973433590494096,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988481556065381,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.30.mlp": [
{
"accuracy": 0.9332266449928284,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9350977912545204,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9425872303545475,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9455157928168774,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9654396362602711,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9684465415775776,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9718980342149734,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9811405893415213,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9830192290246487,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9821723457425833,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9844260476529598,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9908094694837928,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921654621139169,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947911463677883,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952931725420058,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960397034883499,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984408459858969,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.31.self_attn": [
{
"accuracy": 0.9590351842343807,
"total_bits": 89665536,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9609398990869522,
"total_bits": 92221440,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9658170193433762,
"total_bits": 95758848,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9747057911008596,
"total_bits": 112272384,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9792602155357599,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9799246173352003,
"total_bits": 132980224,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64,
"3": 64
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9857713486999273,
"total_bits": 169613312,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.986726806499064,
"total_bits": 169745920,
"q_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880749424919486,
"total_bits": 171195392,
"q_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 64
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884013496339321,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894775114953518,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9904345842078328,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9909303383901715,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9917902136221528,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949478320777416,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958028621040285,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960751314647496,
"total_bits": 253499392,
"q_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977590611670166,
"total_bits": 265838592,
"q_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 32
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989018137566745,
"total_bits": 337385472,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.31.mlp": [
{
"accuracy": 0.9546383582055569,
"total_bits": 395461696,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9560396857559681,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9603356756269932,
"total_bits": 457272320,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9623103924095631,
"total_bits": 512977920,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9766744896769524,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9787986576557159,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9808196611702442,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871868267655373,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884815644472837,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9879759252071381,
"total_bits": 751543392,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9896086370572448,
"total_bits": 766018560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937729584053159,
"total_bits": 924689504,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947134619578719,
"total_bits": 939164672,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.996340551180765,
"total_bits": 1069524064,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967797331046313,
"total_bits": 1110384896,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972505790647119,
"total_bits": 1209999616,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987354436889291,
"total_bits": 1415520512,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.norm.norm": null,
"lm_head.linear": null
},
"strategy": {
"model.layers.0.self_attn": {
"accuracy": 0.9900966100394726,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.0.mlp": {
"accuracy": 0.9714913945645094,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.1.self_attn": {
"accuracy": 0.9886683244258165,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.1.mlp": {
"accuracy": 0.9850738467648625,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.2.self_attn": {
"accuracy": 0.9849622352048755,
"total_bits": 132913152,
"q_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.2.mlp": {
"accuracy": 0.9678649716079235,
"total_bits": 409224256,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
"model.layers.3.self_attn": {
"accuracy": 0.9910929119214416,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.3.mlp": {
"accuracy": 0.9767712149769068,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.4.self_attn": {
"accuracy": 0.9920551627874374,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.4.mlp": {
"accuracy": 0.9729410167783499,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.5.self_attn": {
"accuracy": 0.987149897031486,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.5.mlp": {
"accuracy": 0.9809325225651264,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.6.self_attn": {
"accuracy": 0.9894773187115788,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.6.mlp": {
"accuracy": 0.9808528944849968,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.7.self_attn": {
"accuracy": 0.9874788299202919,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.7.mlp": {
"accuracy": 0.9772119633853436,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.8.self_attn": {
"accuracy": 0.9874300016090274,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.8.mlp": {
"accuracy": 0.9783470071852207,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.9.self_attn": {
"accuracy": 0.9913717247545719,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.9.mlp": {
"accuracy": 0.9775536004453897,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.10.self_attn": {
"accuracy": 0.9907076256349683,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.10.mlp": {
"accuracy": 0.9771080315113068,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.11.self_attn": {
"accuracy": 0.991933386772871,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.11.mlp": {
"accuracy": 0.9771793968975544,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.12.self_attn": {
"accuracy": 0.9914765367284417,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.12.mlp": {
"accuracy": 0.9767845403403044,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.13.self_attn": {
"accuracy": 0.9903721353039145,
"total_bits": 220469248,
"q_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.13.mlp": {
"accuracy": 0.9753897320479155,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.14.self_attn": {
"accuracy": 0.9915214339271188,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.14.mlp": {
"accuracy": 0.9751418121159077,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.15.self_attn": {
"accuracy": 0.9828801974654198,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.15.mlp": {
"accuracy": 0.9758126344531775,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.16.self_attn": {
"accuracy": 0.9922630563378334,
"total_bits": 223535104,
"q_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"6": 32
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.16.mlp": {
"accuracy": 0.976141382008791,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.17.self_attn": {
"accuracy": 0.9845194080844522,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.17.mlp": {
"accuracy": 0.9753408562391996,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.18.self_attn": {
"accuracy": 0.9859072640538216,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.18.mlp": {
"accuracy": 0.9762427378445864,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.19.self_attn": {
"accuracy": 0.9876099163666368,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.19.mlp": {
"accuracy": 0.9778970573097467,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.20.self_attn": {
"accuracy": 0.989013091661036,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.20.mlp": {
"accuracy": 0.9802852533757687,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.21.self_attn": {
"accuracy": 0.9902572967112064,
"total_bits": 181592064,
"q_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.21.mlp": {
"accuracy": 0.9700090549886227,
"total_bits": 637454432,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.22.self_attn": {
"accuracy": 0.9901307374238968,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.22.mlp": {
"accuracy": 0.9825082197785378,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.23.self_attn": {
"accuracy": 0.990400604903698,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.23.mlp": {
"accuracy": 0.9672388434410095,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.24.self_attn": {
"accuracy": 0.9915370875969529,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.24.mlp": {
"accuracy": 0.9832949731498957,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.25.self_attn": {
"accuracy": 0.9906275244429708,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.25.mlp": {
"accuracy": 0.9685135409235954,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.26.self_attn": {
"accuracy": 0.9894789671525359,
"total_bits": 174923264,
"q_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.26.mlp": {
"accuracy": 0.9839572068303823,
"total_bits": 739476480,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.27.self_attn": {
"accuracy": 0.9918317189440131,
"total_bits": 175750144,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.27.mlp": {
"accuracy": 0.9698762781918049,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.28.self_attn": {
"accuracy": 0.9908324750140309,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.28.mlp": {
"accuracy": 0.9820746257901192,
"total_bits": 728741472,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
"model.layers.29.self_attn": {
"accuracy": 0.9868638897314668,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
"model.layers.29.mlp": {
"accuracy": 0.9691478870809078,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.30.self_attn": {
"accuracy": 0.9900705330073833,
"total_bits": 179253248,
"q_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"5": 64
},
"bits": [
5
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"5": 64,
"4": 64
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
}
},
"model.layers.30.mlp": {
"accuracy": 0.9684465415775776,
"total_bits": 592872448,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
"model.layers.31.self_attn": {
"accuracy": 0.9884013496339321,
"total_bits": 173563904,
"q_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
"model.layers.31.mlp": {
"accuracy": 0.9766744896769524,
"total_bits": 578397280,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
}
}
}