{
"measurement": {
"model.layers.0.self_attn": [
{
"accuracy": 0.9633780158075848,
"total_bits": 187058176,
"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": 0.9692641168244576,
"total_bits": 192563200,
"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": 0.9715725616797021,
"total_bits": 197738240,
"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": 0.9787324781676656,
"total_bits": 234305024,
"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": 0.9791449978457469,
"total_bits": 276964352,
"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": 0.981885939983553,
"total_bits": 277163264,
"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": 0.9805063505313898,
"total_bits": 355869696,
"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": 0.9832680005286085,
"total_bits": 356068608,
"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": 0.9823378816639122,
"total_bits": 359029248,
"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": 0.9831473921964827,
"total_bits": 364154880,
"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": 0.9890103819044797,
"total_bits": 365243648,
"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": 0.9908859468506355,
"total_bits": 368204288,
"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": 0.989375744897284,
"total_bits": 373442560,
"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": 0.9916205357171988,
"total_bits": 378966016,
"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": 0.993256408647683,
"total_bits": 458761216,
"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": 0.9956138971212664,
"total_bits": 467046400,
"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": 0.9934863144237744,
"total_bits": 532030464,
"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": 0.9973319692194069,
"total_bits": 552898560,
"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": 0.9980091278802751,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9704362627020792,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9739386376091524,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9760816455573628,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9766874642748582,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.994773114467726,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.994613148847987,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950731560269273,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976783833668665,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978365554967528,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976901326444931,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982024696578983,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986475943848123,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987651172697568,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989316758185037,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989999693766309,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990912582587427,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992835652090818,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.1.self_attn": [
{
"accuracy": 0.9868051656766942,
"total_bits": 187058176,
"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": 0.9886465369674721,
"total_bits": 192563200,
"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": 0.9898352650986788,
"total_bits": 197738240,
"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": 0.9935704116084564,
"total_bits": 234305024,
"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": 0.9939549532504507,
"total_bits": 276964352,
"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": 0.9941838346942524,
"total_bits": 277163264,
"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": 0.9960099553603581,
"total_bits": 355869696,
"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": 0.9961669465287981,
"total_bits": 356068608,
"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": 0.9964304118582635,
"total_bits": 359029248,
"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": 0.9965686079261726,
"total_bits": 364154880,
"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": 0.9970009071081509,
"total_bits": 365243648,
"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": 0.9972035576279986,
"total_bits": 368204288,
"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": 0.9973114639548217,
"total_bits": 373442560,
"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": 0.9975292734299345,
"total_bits": 378966016,
"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": 0.9984792315908156,
"total_bits": 458761216,
"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": 0.9986981410321804,
"total_bits": 467046400,
"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": 0.9988928116409194,
"total_bits": 532030464,
"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": 0.9992711761890525,
"total_bits": 552898560,
"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": 0.9996620267332841,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9568108361410468,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9662030392179364,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.970864892741175,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9708684029939928,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963543473614233,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966693160975794,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966972510569336,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.998008028081753,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977012829530347,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975392582924351,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981112460909093,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.998866756639664,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991046687286335,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.999178268578314,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991887343976965,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991908923624769,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993620324553616,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.2.self_attn": [
{
"accuracy": 0.9986109847025211,
"total_bits": 187058176,
"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": 0.998731068758245,
"total_bits": 192563200,
"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": 0.9988855821705472,
"total_bits": 197738240,
"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": 0.999268414187674,
"total_bits": 234305024,
"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": 0.9993078956965626,
"total_bits": 276964352,
"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": 0.999346013484204,
"total_bits": 277163264,
"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": 0.9995085165321239,
"total_bits": 355869696,
"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": 0.9995530018487969,
"total_bits": 356068608,
"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": 0.9996172552415182,
"total_bits": 359029248,
"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": 0.9996308210668902,
"total_bits": 364154880,
"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": 0.9996631152921749,
"total_bits": 365243648,
"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": 0.9996825103291102,
"total_bits": 368204288,
"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": 0.9996990338955871,
"total_bits": 373442560,
"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": 0.9997203969220496,
"total_bits": 378966016,
"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": 0.9998187247060981,
"total_bits": 458761216,
"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": 0.9998600027699324,
"total_bits": 467046400,
"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": 0.9998579113468479,
"total_bits": 532030464,
"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": 0.9999232440218053,
"total_bits": 552898560,
"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": 0.9999613999995123,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9978763553755063,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.997955334250276,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981783754437378,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982282128859017,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990511053015085,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991310960643827,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.999207863568843,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995110599694798,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.999560493821168,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.999529110702784,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995905251677246,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997619684298499,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997876247070963,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998628419297778,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998776081453176,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998921498611851,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9999632598852546,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.3.self_attn": [
{
"accuracy": 0.9979705182884476,
"total_bits": 187058176,
"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": 0.998140637983421,
"total_bits": 192563200,
"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": 0.9983868864119837,
"total_bits": 197738240,
"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": 0.9988559677608704,
"total_bits": 234305024,
"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": 0.9990254597781275,
"total_bits": 276964352,
"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": 0.9990628788210941,
"total_bits": 277163264,
"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": 0.9993052178148278,
"total_bits": 355869696,
"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": 0.9993523182653773,
"total_bits": 356068608,
"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": 0.9994200314333859,
"total_bits": 359029248,
"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": 0.9994448381537375,
"total_bits": 364154880,
"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": 0.9995121686354794,
"total_bits": 365243648,
"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": 0.9995613476187971,
"total_bits": 368204288,
"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": 0.9995645447925199,
"total_bits": 373442560,
"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": 0.9996175405333481,
"total_bits": 378966016,
"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": 0.9997568366058036,
"total_bits": 458761216,
"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": 0.9998049570452789,
"total_bits": 467046400,
"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": 0.9998120537485718,
"total_bits": 532030464,
"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": 0.9998926118324131,
"total_bits": 552898560,
"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": 0.9999442605114478,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9939329585084986,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941873660458154,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9944403580095815,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9945165416912028,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971576597370011,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965227708759669,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966057022102177,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977495723440754,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968680647557209,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986977341192725,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984474848882344,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991364860171942,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987080191280447,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991987417574579,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996071584219431,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996289042227497,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996743552106445,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.4.self_attn": [
{
"accuracy": 0.996020222928277,
"total_bits": 187058176,
"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": 0.9973327656717677,
"total_bits": 192563200,
"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": 0.9978216522201699,
"total_bits": 197738240,
"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": 0.9984432792315554,
"total_bits": 234305024,
"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": 0.9986197289422547,
"total_bits": 276964352,
"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": 0.9986485499753862,
"total_bits": 277163264,
"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": 0.9990993875330069,
"total_bits": 355869696,
"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": 0.9991241100238097,
"total_bits": 356068608,
"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": 0.9991583328412258,
"total_bits": 359029248,
"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": 0.9991885077079611,
"total_bits": 364154880,
"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": 0.9993164262820133,
"total_bits": 365243648,
"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": 0.9993794973768098,
"total_bits": 368204288,
"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": 0.9994082065256263,
"total_bits": 373442560,
"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": 0.9994707318286313,
"total_bits": 378966016,
"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": 0.9996546701341255,
"total_bits": 458761216,
"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": 0.9997322657367668,
"total_bits": 467046400,
"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": 0.9997302508678618,
"total_bits": 532030464,
"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": 0.9998614091030169,
"total_bits": 552898560,
"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": 0.9999282027510115,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9966817127883827,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967942929973728,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971211091002548,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971935366292631,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984724478808379,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985860419906921,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986905687102606,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992173165971364,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992700190356884,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992364001396039,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.999332502579292,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996109805268996,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996672744776955,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997855993945115,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.999800381687237,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998190843390674,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9999397931360842,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.5.self_attn": [
{
"accuracy": 0.9965810841929755,
"total_bits": 187058176,
"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": 0.9968916223457965,
"total_bits": 192563200,
"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": 0.9974460566269332,
"total_bits": 197738240,
"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": 0.9981023068601069,
"total_bits": 234305024,
"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": 0.998374526278655,
"total_bits": 276964352,
"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": 0.9984232698334381,
"total_bits": 277163264,
"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": 0.9988550277439093,
"total_bits": 355869696,
"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": 0.9989506736593811,
"total_bits": 356068608,
"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": 0.9990630953290843,
"total_bits": 359029248,
"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": 0.9991057480955007,
"total_bits": 364154880,
"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": 0.9991969098306359,
"total_bits": 365243648,
"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": 0.9992672424015933,
"total_bits": 368204288,
"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": 0.9993130095964787,
"total_bits": 373442560,
"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": 0.9993707000486258,
"total_bits": 378966016,
"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": 0.9996133721525486,
"total_bits": 458761216,
"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": 0.9996850255977877,
"total_bits": 467046400,
"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": 0.9997027258934011,
"total_bits": 532030464,
"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": 0.9998337997814432,
"total_bits": 552898560,
"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": 0.9999142192678655,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9960742756283205,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962093727540617,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966147862448308,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.996706453605408,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981564332326678,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983065246824959,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984362583123393,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990687329309845,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991441770260664,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990714217266558,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991900736992982,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995300803627623,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995971695829403,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997381645317593,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997582560179964,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997810682574422,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.999930696541828,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.6.self_attn": [
{
"accuracy": 0.9965270067271041,
"total_bits": 187058176,
"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": 0.9967775266298926,
"total_bits": 192563200,
"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": 0.9972040914244166,
"total_bits": 197738240,
"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": 0.9977791663565624,
"total_bits": 234305024,
"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": 0.9982602675139022,
"total_bits": 276964352,
"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": 0.9983165264252181,
"total_bits": 277163264,
"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": 0.9988782595120076,
"total_bits": 355869696,
"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": 0.9989338600916151,
"total_bits": 356068608,
"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": 0.9990324450742551,
"total_bits": 359029248,
"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": 0.9990763320759135,
"total_bits": 364154880,
"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": 0.9991171474960682,
"total_bits": 365243648,
"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": 0.9991740759854254,
"total_bits": 368204288,
"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": 0.999302826040596,
"total_bits": 373442560,
"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": 0.9993655727012083,
"total_bits": 378966016,
"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": 0.9996074164764172,
"total_bits": 458761216,
"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": 0.9996732345760776,
"total_bits": 467046400,
"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": 0.999703180864392,
"total_bits": 532030464,
"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": 0.9998223930804095,
"total_bits": 552898560,
"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": 0.9999125953690964,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9948893216757202,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950711749077431,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956331746652722,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957566313494585,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.997605532530303,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977645072105684,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979397929824987,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988016445330639,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988047114120012,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987943499857609,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988676480501645,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993823994867087,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994720730755944,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996559182227908,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996872573132063,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997195642484401,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.999910548331651,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.7.self_attn": [
{
"accuracy": 0.9957939597876057,
"total_bits": 187058176,
"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": 0.9961778090678548,
"total_bits": 192563200,
"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": 0.9966265997040624,
"total_bits": 197738240,
"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": 0.9974873938818315,
"total_bits": 234305024,
"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": 0.9979598541731799,
"total_bits": 276964352,
"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": 0.9980206703889722,
"total_bits": 277163264,
"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": 0.9986550384446194,
"total_bits": 355869696,
"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": 0.9987167631717104,
"total_bits": 356068608,
"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": 0.9987948215198948,
"total_bits": 359029248,
"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": 0.9988598046628269,
"total_bits": 364154880,
"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": 0.998982630590418,
"total_bits": 365243648,
"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": 0.9990574026137198,
"total_bits": 368204288,
"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": 0.9991183307260814,
"total_bits": 373442560,
"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": 0.999192279524562,
"total_bits": 378966016,
"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": 0.9995119264104256,
"total_bits": 458761216,
"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": 0.9995893904629874,
"total_bits": 467046400,
"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": 0.9996401708646628,
"total_bits": 532030464,
"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": 0.999775049170957,
"total_bits": 552898560,
"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": 0.9998938560518007,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9939198071431172,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941310447052514,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948330145500797,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949888370465487,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971619092577481,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973868777007354,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976136562242908,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985793537680844,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986897970267915,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985756068704895,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987531732951634,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992768217144705,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993804411144658,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996142648436178,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996262329299689,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996674049578264,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998907581550731,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.8.self_attn": [
{
"accuracy": 0.9952423975459839,
"total_bits": 187058176,
"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": 0.9954791246227136,
"total_bits": 192563200,
"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": 0.9961064460466763,
"total_bits": 197738240,
"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": 0.9972128760265676,
"total_bits": 234305024,
"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": 0.9975925219294272,
"total_bits": 276964352,
"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": 0.9976352298828332,
"total_bits": 277163264,
"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": 0.9984520308452806,
"total_bits": 355869696,
"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": 0.9984962187315288,
"total_bits": 356068608,
"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": 0.9985980485063536,
"total_bits": 359029248,
"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": 0.9986813370334474,
"total_bits": 364154880,
"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": 0.9987966061300156,
"total_bits": 365243648,
"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": 0.9988927175290883,
"total_bits": 368204288,
"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": 0.9989765406169585,
"total_bits": 373442560,
"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": 0.9990652835664438,
"total_bits": 378966016,
"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": 0.9994311953992828,
"total_bits": 458761216,
"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": 0.9995245695080437,
"total_bits": 467046400,
"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": 0.9995921712803753,
"total_bits": 532030464,
"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": 0.9997475392944914,
"total_bits": 552898560,
"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": 0.999883105292515,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9931253293333085,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933516622864103,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.99413753580302,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943142169106164,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968025064537007,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970489648955041,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973124893455717,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983054036274552,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984584306182928,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983552611727071,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985875172056503,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991122443124799,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992748123109586,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994351439113656,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994822459197358,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995217602393639,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996910506804277,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.9.self_attn": [
{
"accuracy": 0.9945872894390241,
"total_bits": 187058176,
"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": 0.9947678235506541,
"total_bits": 192563200,
"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": 0.9955284094903618,
"total_bits": 197738240,
"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": 0.9967399611456418,
"total_bits": 234305024,
"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": 0.9971920607954656,
"total_bits": 276964352,
"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": 0.9971866661952319,
"total_bits": 277163264,
"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": 0.9981834856903573,
"total_bits": 355869696,
"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": 0.9982520875545513,
"total_bits": 356068608,
"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": 0.9984170914762408,
"total_bits": 359029248,
"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": 0.9985011708677599,
"total_bits": 364154880,
"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": 0.9985651867341643,
"total_bits": 365243648,
"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": 0.9986985636152033,
"total_bits": 368204288,
"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": 0.9988083317967769,
"total_bits": 373442560,
"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": 0.9989390080567988,
"total_bits": 378966016,
"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": 0.9993202289486746,
"total_bits": 458761216,
"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": 0.9994480240413012,
"total_bits": 467046400,
"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": 0.9995109837358254,
"total_bits": 532030464,
"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": 0.9997066189067769,
"total_bits": 552898560,
"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": 0.9998594479077024,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.992043939933769,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923191109280053,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9932396076129455,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934434748597836,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962982179866614,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965925060310646,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968951622048687,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981340857369727,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982816189907393,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981396560875797,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983756361516977,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990575841848591,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991881887269157,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994939281651137,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995133531920146,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995683321597569,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998538782318448,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.10.self_attn": [
{
"accuracy": 0.9938754206523299,
"total_bits": 187058176,
"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": 0.9942111311431386,
"total_bits": 192563200,
"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": 0.9948717186129407,
"total_bits": 197738240,
"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": 0.9962090697311061,
"total_bits": 234305024,
"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": 0.99687399425985,
"total_bits": 276964352,
"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": 0.9969597949744448,
"total_bits": 277163264,
"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": 0.9979489965205032,
"total_bits": 355869696,
"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": 0.9980628776155707,
"total_bits": 356068608,
"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": 0.9982410468663531,
"total_bits": 359029248,
"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": 0.9983176982986104,
"total_bits": 364154880,
"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": 0.9984367276468363,
"total_bits": 365243648,
"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": 0.9985512792083778,
"total_bits": 368204288,
"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": 0.9986553810549116,
"total_bits": 373442560,
"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": 0.9987666865957803,
"total_bits": 378966016,
"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": 0.9992413532699606,
"total_bits": 458761216,
"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": 0.9993759552828698,
"total_bits": 467046400,
"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": 0.9994414092108951,
"total_bits": 532030464,
"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": 0.9996590818980047,
"total_bits": 552898560,
"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": 0.9998483979614224,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9914018180907557,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916893903558192,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927396927470047,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9929742175399473,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9959624211524466,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962837378567967,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966284535751727,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979484717141053,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981169556451374,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979873882349286,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982328044287371,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989774234157889,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991213770025704,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994518721436371,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.999475621992085,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995396209461002,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998513585064517,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.11.self_attn": [
{
"accuracy": 0.9928440469513206,
"total_bits": 187058176,
"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": 0.9932965910503346,
"total_bits": 192563200,
"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": 0.9939333088205833,
"total_bits": 197738240,
"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": 0.9954658182901576,
"total_bits": 234305024,
"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": 0.9964896908618117,
"total_bits": 276964352,
"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": 0.9965633928138566,
"total_bits": 277163264,
"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": 0.9977593422364933,
"total_bits": 355869696,
"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": 0.9978591614451847,
"total_bits": 356068608,
"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": 0.9980105745581616,
"total_bits": 359029248,
"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": 0.9980507831113707,
"total_bits": 364154880,
"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": 0.9982445613223765,
"total_bits": 365243648,
"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": 0.9983680912930715,
"total_bits": 368204288,
"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": 0.9984592272238316,
"total_bits": 373442560,
"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": 0.9985818490753636,
"total_bits": 378966016,
"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": 0.9991352553601915,
"total_bits": 458761216,
"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": 0.9992765798325356,
"total_bits": 467046400,
"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": 0.9994066720843732,
"total_bits": 532030464,
"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": 0.9996016272083292,
"total_bits": 552898560,
"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": 0.9998266308539314,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9902466877951825,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9905668137300956,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9917730027319569,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920419882737884,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954022829862017,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.995769037901865,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961588749735567,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977009164500296,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978849797840476,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976915257713316,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979786977453745,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988296675391642,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989958006925391,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993782498594986,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993989235050609,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994703989405503,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998341428573966,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.12.self_attn": [
{
"accuracy": 0.9912717450949314,
"total_bits": 187058176,
"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": 0.9917156993607549,
"total_bits": 192563200,
"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": 0.9925841783222399,
"total_bits": 197738240,
"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": 0.9945243405129173,
"total_bits": 234305024,
"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": 0.9956051895860583,
"total_bits": 276964352,
"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": 0.9956869607619745,
"total_bits": 277163264,
"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": 0.9971920689813009,
"total_bits": 355869696,
"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": 0.9972881252117651,
"total_bits": 356068608,
"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": 0.9975066764133149,
"total_bits": 359029248,
"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": 0.9976631083991379,
"total_bits": 364154880,
"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": 0.997806893747398,
"total_bits": 365243648,
"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": 0.9979558228698903,
"total_bits": 368204288,
"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": 0.9981150537067535,
"total_bits": 373442560,
"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": 0.9982714896261888,
"total_bits": 378966016,
"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": 0.9989485443925092,
"total_bits": 458761216,
"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": 0.9991272758281986,
"total_bits": 467046400,
"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": 0.9992621751275452,
"total_bits": 532030464,
"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": 0.9995028655225484,
"total_bits": 552898560,
"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": 0.999789019912561,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9888443254228485,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9891998870788437,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906266295400105,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9909514874268911,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947322809970692,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951522261958177,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956216001760607,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973623131671431,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975757312679076,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973545496253983,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976836781164533,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986596697647321,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988501106762622,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992875572643243,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993121192704788,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993989320983497,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9998100126333731,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.13.self_attn": [
{
"accuracy": 0.9905007119094463,
"total_bits": 187058176,
"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": 0.9911959503875359,
"total_bits": 192563200,
"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": 0.9920181046779218,
"total_bits": 197738240,
"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": 0.9942268429961252,
"total_bits": 234305024,
"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": 0.9951928927514114,
"total_bits": 276964352,
"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": 0.9953054202794048,
"total_bits": 277163264,
"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": 0.9970759469403052,
"total_bits": 355869696,
"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": 0.9971580239576533,
"total_bits": 356068608,
"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": 0.9974132316113499,
"total_bits": 359029248,
"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": 0.9975316708625638,
"total_bits": 364154880,
"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": 0.9976228520727569,
"total_bits": 365243648,
"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": 0.9977918221051559,
"total_bits": 368204288,
"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": 0.9979670831664024,
"total_bits": 373442560,
"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": 0.9981149643059134,
"total_bits": 378966016,
"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": 0.9988686211260134,
"total_bits": 458761216,
"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": 0.9990521652783316,
"total_bits": 467046400,
"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": 0.9992520444303457,
"total_bits": 532030464,
"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": 0.9994808473811166,
"total_bits": 552898560,
"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": 0.999790265825332,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.987675767489954,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880976230886421,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9896556000274263,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9899980775325706,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9942425447025973,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947004056486645,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952038585355407,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971148146586003,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973488066982674,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971092562681358,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974705013323968,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985325084692848,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987428276738348,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.999217511923665,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992458385838147,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993381905767723,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997904115734855,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.14.self_attn": [
{
"accuracy": 0.9889000405214334,
"total_bits": 187058176,
"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": 0.9893065452771751,
"total_bits": 192563200,
"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": 0.9905108060246628,
"total_bits": 197738240,
"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": 0.9929863993617657,
"total_bits": 234305024,
"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": 0.9942441030013326,
"total_bits": 276964352,
"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": 0.9944190393496108,
"total_bits": 277163264,
"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": 0.996618152020107,
"total_bits": 355869696,
"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": 0.9967589760811901,
"total_bits": 356068608,
"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": 0.9969852532897341,
"total_bits": 359029248,
"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": 0.9971683206355297,
"total_bits": 364154880,
"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": 0.9971596029170445,
"total_bits": 365243648,
"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": 0.9973735002384178,
"total_bits": 368204288,
"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": 0.9976112527809547,
"total_bits": 373442560,
"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": 0.9978046930444083,
"total_bits": 378966016,
"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": 0.9986822212752151,
"total_bits": 458761216,
"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": 0.9988924570988227,
"total_bits": 467046400,
"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": 0.9991262871343153,
"total_bits": 532030464,
"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": 0.999404846859091,
"total_bits": 552898560,
"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": 0.9997553999373371,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9861941168476877,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9866640232013244,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884287231650791,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9888259538596398,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935171274272235,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940348417792273,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946125237222173,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.996741748929612,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970054964069277,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967385980937826,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971474277982978,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983432167393499,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985818044086429,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991159144770599,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991487348329668,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992561289415646,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997635304063902,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.15.self_attn": [
{
"accuracy": 0.9878551785000845,
"total_bits": 187058176,
"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": 0.9882868992556867,
"total_bits": 192563200,
"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": 0.9894773617975021,
"total_bits": 197738240,
"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": 0.9920038341446534,
"total_bits": 234305024,
"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": 0.9938774931636688,
"total_bits": 276964352,
"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": 0.9939658825628852,
"total_bits": 277163264,
"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": 0.9964134375393194,
"total_bits": 355869696,
"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": 0.9964962393216985,
"total_bits": 356068608,
"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": 0.9967469910012656,
"total_bits": 359029248,
"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": 0.996964983127423,
"total_bits": 364154880,
"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": 0.9969480008033937,
"total_bits": 365243648,
"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": 0.9971740201826355,
"total_bits": 368204288,
"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": 0.997399494333781,
"total_bits": 373442560,
"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": 0.9976273491108594,
"total_bits": 378966016,
"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": 0.9985722530584195,
"total_bits": 458761216,
"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": 0.9988005804112426,
"total_bits": 467046400,
"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": 0.9990604842774031,
"total_bits": 532030464,
"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": 0.9993331886086564,
"total_bits": 552898560,
"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": 0.9997327606994842,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.984824665429953,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9853527230259619,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9873087435078464,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9877396748940411,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9929019589173166,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934662593234527,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940987922447292,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964333421452657,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967216954468504,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964331481605768,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.996879765224692,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981898707644034,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984492987982536,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990334898895142,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990703810012507,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991855098633096,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997416017676618,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.16.self_attn": [
{
"accuracy": 0.9861873960808704,
"total_bits": 187058176,
"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": 0.9867696480727509,
"total_bits": 192563200,
"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": 0.9880175937555338,
"total_bits": 197738240,
"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": 0.9906751160313817,
"total_bits": 234305024,
"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": 0.9930040633403941,
"total_bits": 276964352,
"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": 0.9931805365914969,
"total_bits": 277163264,
"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": 0.9958617606960041,
"total_bits": 355869696,
"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": 0.9960312641617891,
"total_bits": 356068608,
"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": 0.9963124069434247,
"total_bits": 359029248,
"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": 0.9965476451662222,
"total_bits": 364154880,
"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": 0.99655794271415,
"total_bits": 365243648,
"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": 0.9968033332965875,
"total_bits": 368204288,
"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": 0.9970194121308037,
"total_bits": 373442560,
"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": 0.9972958034210789,
"total_bits": 378966016,
"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": 0.998388161472241,
"total_bits": 458761216,
"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": 0.9986458615871685,
"total_bits": 467046400,
"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": 0.9989567399852125,
"total_bits": 532030464,
"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": 0.9992467547295076,
"total_bits": 552898560,
"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": 0.9997049128685734,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9834483575664068,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9840285623152005,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9862022839094463,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9866786892770937,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9922639983951261,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9928806384437179,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935740766017452,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961106717512992,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964229134279058,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961156505037492,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966016845581563,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980301622754747,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983109298198926,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989446043465776,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989888316087768,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991134643266713,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9997184992580993,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.17.self_attn": [
{
"accuracy": 0.984402555580202,
"total_bits": 187058176,
"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": 0.9849709782161211,
"total_bits": 192563200,
"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": 0.9864940335483927,
"total_bits": 197738240,
"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": 0.9899998093514066,
"total_bits": 234305024,
"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": 0.9921310764590376,
"total_bits": 276964352,
"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": 0.9922477778684544,
"total_bits": 277163264,
"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": 0.9955650764811588,
"total_bits": 355869696,
"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": 0.9956539379996493,
"total_bits": 356068608,
"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": 0.9959873908297404,
"total_bits": 359029248,
"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": 0.9962188199922246,
"total_bits": 364154880,
"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": 0.9960539214876726,
"total_bits": 365243648,
"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": 0.9963168348430803,
"total_bits": 368204288,
"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": 0.996653772400398,
"total_bits": 373442560,
"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": 0.9969204803561106,
"total_bits": 378966016,
"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": 0.9981575397786832,
"total_bits": 458761216,
"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": 0.9984439743168064,
"total_bits": 467046400,
"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": 0.9988445727718308,
"total_bits": 532030464,
"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": 0.9991663133719397,
"total_bits": 552898560,
"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": 0.9996812876519183,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9820782073626393,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9826969100946659,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850257962059817,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9855380499931542,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916198285002458,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.992282421156568,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930288096350667,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.995781706952441,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961227141933418,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957964730350987,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963179341958541,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978695264404738,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981710943393409,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988595512879425,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989081275600352,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990448331656424,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996953287604571,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.18.self_attn": [
{
"accuracy": 0.9825729074348745,
"total_bits": 187058176,
"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": 0.9831542147225455,
"total_bits": 192563200,
"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": 0.9849135571306473,
"total_bits": 197738240,
"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": 0.9885207064155686,
"total_bits": 234305024,
"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": 0.9910721952389729,
"total_bits": 276964352,
"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": 0.9911715988020756,
"total_bits": 277163264,
"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": 0.9948121684750444,
"total_bits": 355869696,
"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": 0.994947253040185,
"total_bits": 356068608,
"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": 0.9952688855976847,
"total_bits": 359029248,
"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": 0.9956183763784602,
"total_bits": 364154880,
"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": 0.9956039280973767,
"total_bits": 365243648,
"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": 0.9959308811291856,
"total_bits": 368204288,
"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": 0.9962762144124625,
"total_bits": 373442560,
"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": 0.9965789697108496,
"total_bits": 378966016,
"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": 0.9979288407393094,
"total_bits": 458761216,
"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": 0.9982816909537896,
"total_bits": 467046400,
"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": 0.9986536029701758,
"total_bits": 532030464,
"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": 0.9990749317221344,
"total_bits": 552898560,
"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": 0.9996354736871772,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9807397649182301,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9814064050779531,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9838924655984891,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9844502908618826,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.990977235824654,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916931483159331,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924948567671603,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954378064231653,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958091452472696,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954720421718728,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960338965695548,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977047460607106,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980287837710133,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987662297973753,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988249784127172,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989722958845576,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996692368141865,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.19.self_attn": [
{
"accuracy": 0.98030314403341,
"total_bits": 187058176,
"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": 0.9809748070608628,
"total_bits": 192563200,
"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": 0.9829432452097535,
"total_bits": 197738240,
"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": 0.9865422781654879,
"total_bits": 234305024,
"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": 0.9898240781438193,
"total_bits": 276964352,
"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": 0.989846831996386,
"total_bits": 277163264,
"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": 0.9939272818155587,
"total_bits": 355869696,
"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": 0.9940319478658861,
"total_bits": 356068608,
"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": 0.9946385607809612,
"total_bits": 359029248,
"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": 0.994970180437361,
"total_bits": 364154880,
"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": 0.9948975757426141,
"total_bits": 365243648,
"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": 0.9953407404919792,
"total_bits": 368204288,
"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": 0.9957683820056876,
"total_bits": 373442560,
"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": 0.9961056982681743,
"total_bits": 378966016,
"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": 0.9976585844139519,
"total_bits": 458761216,
"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": 0.9980482566354208,
"total_bits": 467046400,
"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": 0.9984197275480255,
"total_bits": 532030464,
"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": 0.9989337086689806,
"total_bits": 552898560,
"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": 0.9995658870933479,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9793642734814632,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9800719923682903,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9827609941069233,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833584624391637,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9902965356840899,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910674286320021,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919214415069866,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951035122043992,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.995497336378321,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951249828660175,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.995729825778031,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975269291991073,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.997877068590912,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986734739504755,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987328080154073,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9988886332476364,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996458270072349,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.20.self_attn": [
{
"accuracy": 0.9797716590722925,
"total_bits": 187058176,
"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": 0.980771108285377,
"total_bits": 192563200,
"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": 0.982928940977313,
"total_bits": 197738240,
"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": 0.9868936374979583,
"total_bits": 234305024,
"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": 0.9898817215212866,
"total_bits": 276964352,
"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": 0.9900243932773408,
"total_bits": 277163264,
"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": 0.9939379705067136,
"total_bits": 355869696,
"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": 0.9941180906209507,
"total_bits": 356068608,
"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": 0.9944878124729976,
"total_bits": 359029248,
"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": 0.9948465164236137,
"total_bits": 364154880,
"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": 0.9949357796607441,
"total_bits": 365243648,
"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": 0.995284049612421,
"total_bits": 368204288,
"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": 0.9957662362629842,
"total_bits": 373442560,
"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": 0.9960794090293348,
"total_bits": 378966016,
"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": 0.997680507827011,
"total_bits": 458761216,
"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": 0.9980456546304355,
"total_bits": 467046400,
"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": 0.9984347005558544,
"total_bits": 532030464,
"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": 0.998890653920458,
"total_bits": 552898560,
"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": 0.9995675246831716,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9775425439798519,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9783182367682457,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9812701594966807,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9819303412284506,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894373686494011,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.990283185836712,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912152201868594,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946536684506818,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950805702369268,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946934526148987,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953544015726564,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973095684728929,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976896770776397,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985484890455968,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986229598044271,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9987926158845719,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9996090818711213,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.21.self_attn": [
{
"accuracy": 0.9765283798700884,
"total_bits": 187058176,
"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": 0.9774354018859173,
"total_bits": 192563200,
"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": 0.9797796445751661,
"total_bits": 197738240,
"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": 0.9844290788138383,
"total_bits": 234305024,
"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": 0.9879917066152158,
"total_bits": 276964352,
"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": 0.9882162252048913,
"total_bits": 277163264,
"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": 0.9928586192576117,
"total_bits": 355869696,
"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": 0.9931195479523587,
"total_bits": 356068608,
"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": 0.9935847859780648,
"total_bits": 359029248,
"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": 0.9939618553280046,
"total_bits": 364154880,
"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": 0.9940590244031658,
"total_bits": 365243648,
"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": 0.9944872820112658,
"total_bits": 368204288,
"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": 0.9949902207958267,
"total_bits": 373442560,
"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": 0.9954134227928558,
"total_bits": 378966016,
"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": 0.997203994027682,
"total_bits": 458761216,
"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": 0.9976937623853844,
"total_bits": 467046400,
"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": 0.9981462739493796,
"total_bits": 532030464,
"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": 0.9987031561996493,
"total_bits": 552898560,
"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": 0.9994838515076002,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9754537182222855,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9762879284588915,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9794434474683121,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9801582469555893,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884768669542513,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.989393645162253,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9903914209602303,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941668308732149,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.994633373681848,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9942146130886517,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949280085826391,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970661114876795,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974778554598359,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984208584309703,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984978627793393,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986811023340315,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995748275817421,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.22.self_attn": [
{
"accuracy": 0.9745552480024727,
"total_bits": 187058176,
"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": 0.9755557159844198,
"total_bits": 192563200,
"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": 0.9782381763583735,
"total_bits": 197738240,
"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": 0.9838098178274537,
"total_bits": 234305024,
"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": 0.986977079491082,
"total_bits": 276964352,
"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": 0.9872301794882667,
"total_bits": 277163264,
"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": 0.9924089763754684,
"total_bits": 355869696,
"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": 0.9926247257543238,
"total_bits": 356068608,
"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": 0.9931480333356089,
"total_bits": 359029248,
"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": 0.9935610792109448,
"total_bits": 364154880,
"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": 0.9935019896610787,
"total_bits": 365243648,
"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": 0.9940149593902262,
"total_bits": 368204288,
"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": 0.9946233309119156,
"total_bits": 373442560,
"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": 0.9950477395598826,
"total_bits": 378966016,
"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": 0.9970358245099258,
"total_bits": 458761216,
"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": 0.9975314945607495,
"total_bits": 467046400,
"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": 0.9980253310763816,
"total_bits": 532030464,
"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": 0.9986633783892581,
"total_bits": 552898560,
"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": 0.9994708992399912,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9746878580435326,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9755706978275588,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.978847233186427,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9795762969573077,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9881814562862641,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9891142678986254,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9901197601943031,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940411955804417,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9945067391289693,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940632533674177,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947982473347924,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969890490419379,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974131364816505,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983829146398133,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984574232703859,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9986412867192963,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995668564224616,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.23.self_attn": [
{
"accuracy": 0.9791547464402882,
"total_bits": 187058176,
"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": 0.9799550537902274,
"total_bits": 192563200,
"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": 0.9820938670615617,
"total_bits": 197738240,
"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": 0.9861729490619741,
"total_bits": 234305024,
"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": 0.9891797034069896,
"total_bits": 276964352,
"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": 0.9893358205200026,
"total_bits": 277163264,
"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": 0.9934104698416042,
"total_bits": 355869696,
"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": 0.9935805128250075,
"total_bits": 356068608,
"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": 0.9942187085061481,
"total_bits": 359029248,
"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": 0.9945120719555569,
"total_bits": 364154880,
"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": 0.9947210463548177,
"total_bits": 365243648,
"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": 0.9950903625913748,
"total_bits": 368204288,
"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": 0.9955386155603552,
"total_bits": 373442560,
"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": 0.995892478079584,
"total_bits": 378966016,
"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": 0.9975188581125909,
"total_bits": 458761216,
"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": 0.9979184440913071,
"total_bits": 467046400,
"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": 0.9982680534625328,
"total_bits": 532030464,
"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": 0.9988590781576931,
"total_bits": 552898560,
"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": 0.9995201329446374,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.974413168469542,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9752721470829687,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9784373017518144,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.979151922964344,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9878918797169861,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9888626599292222,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9898577649636489,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.993908121863282,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943841054153285,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9939099766657149,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946701296507136,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.99691284608439,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973505306866413,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983438499835565,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9984158722708296,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985956812681827,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995518444216271,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.24.self_attn": [
{
"accuracy": 0.9785734063112422,
"total_bits": 187058176,
"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": 0.9802615822439915,
"total_bits": 192563200,
"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": 0.981864031061138,
"total_bits": 197738240,
"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": 0.9860304558747693,
"total_bits": 234305024,
"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": 0.9894709623614816,
"total_bits": 276964352,
"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": 0.9897990941658223,
"total_bits": 277163264,
"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": 0.9936370299638886,
"total_bits": 355869696,
"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": 0.9940039609362813,
"total_bits": 356068608,
"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": 0.9943780431916055,
"total_bits": 359029248,
"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": 0.9947469036192879,
"total_bits": 364154880,
"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": 0.9949255666058314,
"total_bits": 365243648,
"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": 0.9952721582284492,
"total_bits": 368204288,
"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": 0.9956133059520078,
"total_bits": 373442560,
"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": 0.9959564475881818,
"total_bits": 378966016,
"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": 0.9975619433190379,
"total_bits": 458761216,
"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": 0.9979686888951906,
"total_bits": 467046400,
"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": 0.9983712650259564,
"total_bits": 532030464,
"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": 0.9988685122776875,
"total_bits": 552898560,
"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": 0.9995473155071714,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.973361722928913,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9742302936746886,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9773967178832543,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9780995240925174,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9873247533350399,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9883521304122711,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.989332732548447,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.993639086005523,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941404785617793,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936114424153378,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9944120992563272,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967537239733103,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972213146423823,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982675914591375,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983334995156742,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9985101137334775,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995366068370027,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.25.self_attn": [
{
"accuracy": 0.9806247330025623,
"total_bits": 187058176,
"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": 0.9813208530509943,
"total_bits": 192563200,
"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": 0.9834373001205293,
"total_bits": 197738240,
"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": 0.9871796160740288,
"total_bits": 234305024,
"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": 0.9901653092942739,
"total_bits": 276964352,
"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": 0.9902935433691662,
"total_bits": 277163264,
"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": 0.9939682348876407,
"total_bits": 355869696,
"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": 0.9941358486631591,
"total_bits": 356068608,
"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": 0.9947247700520644,
"total_bits": 359029248,
"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": 0.9949914887792578,
"total_bits": 364154880,
"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": 0.995047857911375,
"total_bits": 365243648,
"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": 0.995458813469955,
"total_bits": 368204288,
"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": 0.9958717450622077,
"total_bits": 373442560,
"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": 0.996273010417721,
"total_bits": 378966016,
"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": 0.9977071860216951,
"total_bits": 458761216,
"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": 0.9981185208185037,
"total_bits": 467046400,
"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": 0.9984215040887264,
"total_bits": 532030464,
"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": 0.9989559821551666,
"total_bits": 552898560,
"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": 0.9995536382377491,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9725105800714932,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9734136852386751,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9766433004682001,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9773552377561205,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9869369962987932,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9879819352767969,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9889845265761802,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934338795903482,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9939481305252564,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934212762225223,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9942404663827467,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966614688991716,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971339232043216,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982167289494291,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982867787016163,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.998468302600821,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9995199589695978,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.26.self_attn": [
{
"accuracy": 0.9782770449589742,
"total_bits": 187058176,
"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": 0.9788017505289692,
"total_bits": 192563200,
"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": 0.9817821484941401,
"total_bits": 197738240,
"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": 0.9856478262104487,
"total_bits": 234305024,
"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": 0.9885916173163998,
"total_bits": 276964352,
"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": 0.9889293205189077,
"total_bits": 277163264,
"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": 0.9929131364792978,
"total_bits": 355869696,
"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": 0.99313963030612,
"total_bits": 356068608,
"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": 0.9941931399761846,
"total_bits": 359029248,
"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": 0.9945573835682712,
"total_bits": 364154880,
"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": 0.9946727480273694,
"total_bits": 365243648,
"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": 0.9950722019972378,
"total_bits": 368204288,
"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": 0.9955098650565273,
"total_bits": 373442560,
"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": 0.9958293691375538,
"total_bits": 378966016,
"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": 0.9974702116175506,
"total_bits": 458761216,
"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": 0.9979149659997538,
"total_bits": 467046400,
"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": 0.9982627485878766,
"total_bits": 532030464,
"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": 0.9988447441872651,
"total_bits": 552898560,
"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": 0.9995119307040066,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9707463170940939,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9716864076295966,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9751126287799132,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9758707364334872,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9860931651078557,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871912812323946,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9882587109643378,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930026033452075,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935537082339195,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930008632179937,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9938622212625647,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964492864857771,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969478134237426,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981076514997863,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981800638151502,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983781170173499,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994928725814374,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.27.self_attn": [
{
"accuracy": 0.9799080564101276,
"total_bits": 187058176,
"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": 0.9806471249382747,
"total_bits": 192563200,
"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": 0.982810198525457,
"total_bits": 197738240,
"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": 0.9861632811983949,
"total_bits": 234305024,
"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": 0.9895845812892443,
"total_bits": 276964352,
"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": 0.9896477090292856,
"total_bits": 277163264,
"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": 0.9936371522857562,
"total_bits": 355869696,
"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": 0.9934367404662465,
"total_bits": 356068608,
"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": 0.9943046903874921,
"total_bits": 359029248,
"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": 0.9945511240021963,
"total_bits": 364154880,
"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": 0.9947543651674335,
"total_bits": 365243648,
"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": 0.9951821831177527,
"total_bits": 368204288,
"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": 0.9956226498868904,
"total_bits": 373442560,
"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": 0.9959815210589257,
"total_bits": 378966016,
"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": 0.9975416207943406,
"total_bits": 458761216,
"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": 0.9979668670505482,
"total_bits": 467046400,
"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": 0.9982651002754114,
"total_bits": 532030464,
"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": 0.9988657294153725,
"total_bits": 552898560,
"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": 0.9995125001740618,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9696155232622435,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9705913573699562,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9741236193987884,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.974918441454831,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9855612743841974,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9867109771523821,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9878426331438517,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927282642426067,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933075595058893,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927409534636689,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936334106973127,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.996319017005398,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968345357154152,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980319814977089,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981141660705601,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9983228671572808,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994685672863852,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.28.self_attn": [
{
"accuracy": 0.9760475293978265,
"total_bits": 187058176,
"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": 0.9772633810674674,
"total_bits": 192563200,
"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": 0.9800652749346275,
"total_bits": 197738240,
"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": 0.9843363083997055,
"total_bits": 234305024,
"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": 0.9875869474030639,
"total_bits": 276964352,
"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": 0.9876585487080248,
"total_bits": 277163264,
"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": 0.9922573260815912,
"total_bits": 355869696,
"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": 0.9924534491116279,
"total_bits": 356068608,
"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": 0.9931191292983529,
"total_bits": 359029248,
"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": 0.9936524190155691,
"total_bits": 364154880,
"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": 0.9940371152234116,
"total_bits": 365243648,
"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": 0.9944395197702474,
"total_bits": 368204288,
"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": 0.9950332326790023,
"total_bits": 373442560,
"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": 0.995362020453046,
"total_bits": 378966016,
"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": 0.9972194662305379,
"total_bits": 458761216,
"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": 0.9976755970427277,
"total_bits": 467046400,
"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": 0.9980400985682776,
"total_bits": 532030464,
"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": 0.9986867315417743,
"total_bits": 552898560,
"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": 0.9994488249633976,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9682956106568638,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.969318523708927,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9731659226511654,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9740627308033014,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.984914685305404,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9861394437520128,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9874169944638485,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923665930124882,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9929980971409302,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924308020670555,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933636140470442,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961615868500973,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966995909712032,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979399961333624,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980370088198566,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982714714470172,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994446970966621,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.29.self_attn": [
{
"accuracy": 0.974849261637581,
"total_bits": 187058176,
"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": 0.9759981915060627,
"total_bits": 192563200,
"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": 0.9792253648077971,
"total_bits": 197738240,
"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": 0.9841352299738088,
"total_bits": 234305024,
"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": 0.9872925775615793,
"total_bits": 276964352,
"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": 0.9876230966397807,
"total_bits": 277163264,
"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": 0.992224821918889,
"total_bits": 355869696,
"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": 0.9924872512684056,
"total_bits": 356068608,
"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": 0.9929462325220045,
"total_bits": 359029248,
"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": 0.9933134680123705,
"total_bits": 364154880,
"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": 0.9936685998507432,
"total_bits": 365243648,
"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": 0.9940780419488683,
"total_bits": 368204288,
"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": 0.9947813588782752,
"total_bits": 373442560,
"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": 0.9951508241877156,
"total_bits": 378966016,
"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": 0.997089593052423,
"total_bits": 458761216,
"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": 0.997560621594618,
"total_bits": 467046400,
"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": 0.9979310286304864,
"total_bits": 532030464,
"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": 0.9986555943492261,
"total_bits": 552898560,
"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": 0.9994273569701411,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9686012620008305,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.969554811029842,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9733309602658999,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9742191027065641,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850233734438294,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9862080127196876,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.987461704466688,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.992429167693971,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9930440118340286,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924683306170138,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933935767529827,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961833552898545,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967147124146944,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979570823978927,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980463899953879,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9982750057855523,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994466993559495,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.30.self_attn": [
{
"accuracy": 0.980396813958099,
"total_bits": 187058176,
"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": 0.9813208503550604,
"total_bits": 192563200,
"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": 0.9836371403775717,
"total_bits": 197738240,
"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": 0.9872346314062413,
"total_bits": 234305024,
"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": 0.9898530338449698,
"total_bits": 276964352,
"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": 0.9900167512854463,
"total_bits": 277163264,
"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": 0.9935989525953406,
"total_bits": 355869696,
"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": 0.9938359689702722,
"total_bits": 356068608,
"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": 0.9942835962497875,
"total_bits": 359029248,
"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": 0.99459583531624,
"total_bits": 364154880,
"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": 0.9949071032093152,
"total_bits": 365243648,
"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": 0.9952706221834218,
"total_bits": 368204288,
"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": 0.995782544885419,
"total_bits": 373442560,
"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": 0.9960975108276072,
"total_bits": 378966016,
"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": 0.9976453968801683,
"total_bits": 458761216,
"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": 0.9980341630411873,
"total_bits": 467046400,
"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": 0.9983166839290214,
"total_bits": 532030464,
"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": 0.9989107642261508,
"total_bits": 552898560,
"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": 0.999513480443152,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9674970529189235,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9685116947481507,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9723359810090378,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9732359341884914,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9844466182648351,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9856835114524553,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9869431058238995,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921485779218768,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927847653529361,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921770940142635,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9931432986877075,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960354548805442,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965909777307197,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978796886157637,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979714237361852,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981988605393685,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994271398019535,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.31.self_attn": [
{
"accuracy": 0.9786801370547006,
"total_bits": 187058176,
"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": 0.9794729332880754,
"total_bits": 192563200,
"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": 0.9823121792195659,
"total_bits": 197738240,
"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": 0.9861081315596637,
"total_bits": 234305024,
"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": 0.9889770312675912,
"total_bits": 276964352,
"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": 0.9893170638058922,
"total_bits": 277163264,
"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": 0.9927511552201682,
"total_bits": 355869696,
"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": 0.9932524482964685,
"total_bits": 356068608,
"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": 0.9937630907730445,
"total_bits": 359029248,
"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": 0.9940731725655496,
"total_bits": 364154880,
"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": 0.994564226853024,
"total_bits": 365243648,
"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": 0.994998653076197,
"total_bits": 368204288,
"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": 0.9954380137404721,
"total_bits": 373442560,
"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": 0.995803409954533,
"total_bits": 378966016,
"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": 0.9974062814634588,
"total_bits": 458761216,
"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": 0.9978761060200142,
"total_bits": 467046400,
"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": 0.998064506566152,
"total_bits": 532030464,
"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": 0.9988666122080758,
"total_bits": 552898560,
"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": 0.9994738713017126,
"total_bits": 708191232,
"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.block_sparse_moe": [
{
"accuracy": 0.9664879371657183,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9675199025751728,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9713035461149717,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9721775437264066,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9839447306861219,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.985223298647294,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9864457128569484,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919078788524004,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925620507105792,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919268002635554,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9929220499715915,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9959090055260611,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964822531726799,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978174612992198,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979076788233789,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9981293681776151,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994142784884101,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.32.self_attn": [
{
"accuracy": 0.9796198815793583,
"total_bits": 187058176,
"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": 0.980486639334183,
"total_bits": 192563200,
"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": 0.9829723427660371,
"total_bits": 197738240,
"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": 0.9869982115679273,
"total_bits": 234305024,
"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": 0.9897251787203315,
"total_bits": 276964352,
"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": 0.9898908571095059,
"total_bits": 277163264,
"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": 0.9936392404825279,
"total_bits": 355869696,
"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": 0.9938632179247705,
"total_bits": 356068608,
"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": 0.9943347756848916,
"total_bits": 359029248,
"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": 0.9946294628500351,
"total_bits": 364154880,
"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": 0.9948698682710528,
"total_bits": 365243648,
"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": 0.995215840294565,
"total_bits": 368204288,
"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": 0.995678867715852,
"total_bits": 373442560,
"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": 0.9960297757367554,
"total_bits": 378966016,
"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": 0.9976110687589665,
"total_bits": 458761216,
"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": 0.9979965128372178,
"total_bits": 467046400,
"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": 0.9983225917921549,
"total_bits": 532030464,
"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": 0.9988640281616857,
"total_bits": 552898560,
"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": 0.9995233721396959,
"total_bits": 708191232,
"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.32.block_sparse_moe": [
{
"accuracy": 0.965767735027169,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9668189611678061,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9705692501248497,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9714477604353114,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9835558061635024,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9848565143091899,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9860683336952015,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9917204793003437,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923852545111195,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9917316497793716,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927470500973102,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958133866005626,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963966521900147,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977701516319556,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978586107823312,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980755102158966,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9994010606284351,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.33.self_attn": [
{
"accuracy": 0.9789281728628435,
"total_bits": 187058176,
"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": 0.9804875327176169,
"total_bits": 192563200,
"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": 0.9828474959358573,
"total_bits": 197738240,
"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": 0.9863290813212332,
"total_bits": 234305024,
"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": 0.9896740348470446,
"total_bits": 276964352,
"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": 0.9899098559919941,
"total_bits": 277163264,
"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": 0.9933366965208399,
"total_bits": 355869696,
"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": 0.9936042648347977,
"total_bits": 356068608,
"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": 0.9940476929523835,
"total_bits": 359029248,
"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": 0.9944034191828809,
"total_bits": 364154880,
"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": 0.9948533887754342,
"total_bits": 365243648,
"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": 0.9952178421479306,
"total_bits": 368204288,
"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": 0.9956425391995397,
"total_bits": 373442560,
"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": 0.9959913481296482,
"total_bits": 378966016,
"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": 0.9975628608555877,
"total_bits": 458761216,
"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": 0.9979571250365361,
"total_bits": 467046400,
"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": 0.9982525251352375,
"total_bits": 532030464,
"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": 0.9988259196030221,
"total_bits": 552898560,
"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": 0.9994975622608572,
"total_bits": 708191232,
"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.33.block_sparse_moe": [
{
"accuracy": 0.9649931444345337,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9660635930731108,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9698722625249311,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9707570418127274,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9831726738790932,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9845208970241641,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9857307246052905,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915334368369689,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9922218004878807,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915320307604576,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925812383084313,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957068140593994,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963144849079024,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977177698165178,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978033928384417,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980174064097044,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.999390004290035,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.34.self_attn": [
{
"accuracy": 0.9777466665071092,
"total_bits": 187058176,
"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": 0.9798036530908001,
"total_bits": 192563200,
"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": 0.9822950823428599,
"total_bits": 197738240,
"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": 0.9867574545358749,
"total_bits": 234305024,
"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": 0.9892917507279076,
"total_bits": 276964352,
"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": 0.989488906668205,
"total_bits": 277163264,
"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": 0.9933413846761381,
"total_bits": 355869696,
"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": 0.9935850355234978,
"total_bits": 356068608,
"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": 0.9941454129293561,
"total_bits": 359029248,
"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": 0.994480329320619,
"total_bits": 364154880,
"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": 0.9947778362241623,
"total_bits": 365243648,
"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": 0.995118102193565,
"total_bits": 368204288,
"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": 0.9955638036818096,
"total_bits": 373442560,
"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": 0.9959022926632315,
"total_bits": 378966016,
"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": 0.9975357185436511,
"total_bits": 458761216,
"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": 0.9979351462771822,
"total_bits": 467046400,
"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": 0.9982647138439413,
"total_bits": 532030464,
"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": 0.9988532723785427,
"total_bits": 552898560,
"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": 0.9995088163659124,
"total_bits": 708191232,
"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.34.block_sparse_moe": [
{
"accuracy": 0.9655536186734313,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9665957544195025,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9702372734287852,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9710842310696056,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9834115434634058,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9847400624206976,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9858931381696541,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916629975563601,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923388287179956,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916600849675504,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9926872682806692,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957745412217551,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963663388288727,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977543143077606,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978381562426588,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980404760541493,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993966884901869,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.35.self_attn": [
{
"accuracy": 0.9850480471688666,
"total_bits": 187058176,
"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": 0.9851736645949515,
"total_bits": 192563200,
"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": 0.9879812013455912,
"total_bits": 197738240,
"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": 0.9901043515358317,
"total_bits": 234305024,
"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": 0.99202587651579,
"total_bits": 276964352,
"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": 0.9924252920195853,
"total_bits": 277163264,
"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": 0.9953526425739064,
"total_bits": 355869696,
"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": 0.9955612673840829,
"total_bits": 356068608,
"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": 0.9959326726507003,
"total_bits": 359029248,
"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": 0.996098882432929,
"total_bits": 364154880,
"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": 0.9959450189980041,
"total_bits": 365243648,
"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": 0.996245580382253,
"total_bits": 368204288,
"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": 0.996956925753406,
"total_bits": 373442560,
"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": 0.9972304237859422,
"total_bits": 378966016,
"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": 0.9983321837284357,
"total_bits": 458761216,
"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": 0.9985984753431311,
"total_bits": 467046400,
"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": 0.9987621191080267,
"total_bits": 532030464,
"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": 0.9992825927390521,
"total_bits": 552898560,
"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": 0.9996315906112279,
"total_bits": 708191232,
"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.35.block_sparse_moe": [
{
"accuracy": 0.964409745249309,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9654991209114853,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9691969375861319,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9700521049334815,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9828505387137595,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9842213158936877,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9853817930837211,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9913774732696382,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920775724672958,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9913749215683263,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9924381745636071,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956291899036028,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962428744000039,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976770200974945,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977625941575848,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979678151246748,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993776930143787,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.36.self_attn": [
{
"accuracy": 0.9821430157968088,
"total_bits": 187058176,
"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": 0.9828577741098247,
"total_bits": 192563200,
"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": 0.9847191370729553,
"total_bits": 197738240,
"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": 0.9880391313988519,
"total_bits": 234305024,
"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": 0.9906159385940746,
"total_bits": 276964352,
"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": 0.9907964362370733,
"total_bits": 277163264,
"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": 0.9943039438711774,
"total_bits": 355869696,
"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": 0.9945600475307161,
"total_bits": 356068608,
"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": 0.9949911683185124,
"total_bits": 359029248,
"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": 0.9951892926014567,
"total_bits": 364154880,
"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": 0.995219915235219,
"total_bits": 365243648,
"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": 0.9955931840905625,
"total_bits": 368204288,
"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": 0.9961543974442113,
"total_bits": 373442560,
"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": 0.9964751261163896,
"total_bits": 378966016,
"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": 0.9978684039700678,
"total_bits": 458761216,
"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": 0.9982154840101046,
"total_bits": 467046400,
"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": 0.9985041620185304,
"total_bits": 532030464,
"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": 0.9990070877431304,
"total_bits": 552898560,
"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": 0.9995699108173874,
"total_bits": 708191232,
"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.36.block_sparse_moe": [
{
"accuracy": 0.9631280264768162,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9644314236938953,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9680641852319241,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9689056609610194,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9825642578990051,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9839481031031985,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.985106611820428,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912308524607828,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919334897330325,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912547994356015,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923304235072512,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955708243222418,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961887413243714,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976438459891238,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977327108959129,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979382997354198,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993687953989904,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.37.self_attn": [
{
"accuracy": 0.9783325276190513,
"total_bits": 187058176,
"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": 0.9792233886393277,
"total_bits": 192563200,
"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": 0.9811778725860151,
"total_bits": 197738240,
"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": 0.9856812704452559,
"total_bits": 234305024,
"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": 0.9887832451954877,
"total_bits": 276964352,
"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": 0.9890571741485282,
"total_bits": 277163264,
"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": 0.9930769733006233,
"total_bits": 355869696,
"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": 0.993402933211703,
"total_bits": 356068608,
"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": 0.9939206118548387,
"total_bits": 359029248,
"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": 0.9942706804722548,
"total_bits": 364154880,
"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": 0.9943301530162755,
"total_bits": 365243648,
"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": 0.99468825952346,
"total_bits": 368204288,
"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": 0.995227002857351,
"total_bits": 373442560,
"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": 0.9956497897541052,
"total_bits": 378966016,
"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": 0.9973599669940182,
"total_bits": 458761216,
"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": 0.9977987601826164,
"total_bits": 467046400,
"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": 0.9981585516974232,
"total_bits": 532030464,
"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": 0.9987576589460092,
"total_bits": 552898560,
"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": 0.9994740463612528,
"total_bits": 708191232,
"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.37.block_sparse_moe": [
{
"accuracy": 0.9635363211365122,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.964710984771189,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9683424638290155,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9691841329300874,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9824952004281314,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9838811570876523,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850219423686596,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.991196106580135,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918810541153347,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912010919253685,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.992285410726541,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955434155356335,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961679398547858,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976312128802467,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977193564103034,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979202330333033,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993637444797634,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.38.self_attn": [
{
"accuracy": 0.9840628102323726,
"total_bits": 187058176,
"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": 0.9849754906210461,
"total_bits": 192563200,
"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": 0.9873822938258711,
"total_bits": 197738240,
"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": 0.9896797572360619,
"total_bits": 234305024,
"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": 0.9917179100999707,
"total_bits": 276964352,
"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": 0.9919832839483493,
"total_bits": 277163264,
"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": 0.9949157149530947,
"total_bits": 355869696,
"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": 0.995268791252257,
"total_bits": 356068608,
"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": 0.9956483822561016,
"total_bits": 359029248,
"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": 0.9958229339804108,
"total_bits": 364154880,
"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": 0.9958192546312746,
"total_bits": 365243648,
"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": 0.996070543414374,
"total_bits": 368204288,
"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": 0.9967871970423546,
"total_bits": 373442560,
"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": 0.9971325117359428,
"total_bits": 378966016,
"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": 0.9982427826188015,
"total_bits": 458761216,
"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": 0.9985610144396656,
"total_bits": 467046400,
"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": 0.9986416941055244,
"total_bits": 532030464,
"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": 0.9992077576967667,
"total_bits": 552898560,
"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": 0.9996134919331924,
"total_bits": 708191232,
"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.38.block_sparse_moe": [
{
"accuracy": 0.9643347184909017,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.965426868219909,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9689401798930607,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9697556779474804,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9827138584499296,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9841102028853799,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9851926381111538,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.991300178152558,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920062895766214,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9912993005525909,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923745086906772,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955928693648035,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962118095105612,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976505282347238,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977440638896568,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.99793304085707,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993688033336137,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.39.self_attn": [
{
"accuracy": 0.9816540715921867,
"total_bits": 187058176,
"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": 0.9826675060352212,
"total_bits": 192563200,
"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": 0.9853483941406012,
"total_bits": 197738240,
"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": 0.9885618601876655,
"total_bits": 234305024,
"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": 0.99054833021211,
"total_bits": 276964352,
"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": 0.9908602056583684,
"total_bits": 277163264,
"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": 0.9942918664187586,
"total_bits": 355869696,
"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": 0.9945732289189962,
"total_bits": 356068608,
"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": 0.9950097359251231,
"total_bits": 359029248,
"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": 0.9952389598015303,
"total_bits": 364154880,
"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": 0.9953172469609662,
"total_bits": 365243648,
"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": 0.995590351227867,
"total_bits": 368204288,
"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": 0.9963468508917446,
"total_bits": 373442560,
"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": 0.9966424929259933,
"total_bits": 378966016,
"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": 0.9979746790149396,
"total_bits": 458761216,
"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": 0.9982992318722321,
"total_bits": 467046400,
"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": 0.9985007359156091,
"total_bits": 532030464,
"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": 0.999040265083558,
"total_bits": 552898560,
"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": 0.999558122736696,
"total_bits": 708191232,
"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.39.block_sparse_moe": [
{
"accuracy": 0.9638478541256565,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9649416570992846,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.968457901262139,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9692786616812411,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9824562783500081,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9838739482587889,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9849613718933573,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9911759908703205,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918938169844056,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9911724588539648,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9922605087049305,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955269987204749,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961536618219199,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976232037389356,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977112127664057,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978991740407717,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993627348203996,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.40.self_attn": [
{
"accuracy": 0.9841205658587185,
"total_bits": 187058176,
"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": 0.9849975313010969,
"total_bits": 192563200,
"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": 0.986854489230992,
"total_bits": 197738240,
"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": 0.9895629793846685,
"total_bits": 234305024,
"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": 0.9917467239214793,
"total_bits": 276964352,
"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": 0.9919874423771704,
"total_bits": 277163264,
"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": 0.9947274528983,
"total_bits": 355869696,
"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": 0.9949960569777575,
"total_bits": 356068608,
"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": 0.9953728346521721,
"total_bits": 359029248,
"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": 0.9955975883885434,
"total_bits": 364154880,
"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": 0.9957772314205373,
"total_bits": 365243648,
"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": 0.9960481801413392,
"total_bits": 368204288,
"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": 0.9966230157787275,
"total_bits": 373442560,
"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": 0.9969403384326908,
"total_bits": 378966016,
"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": 0.998122462151131,
"total_bits": 458761216,
"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": 0.998453784167252,
"total_bits": 467046400,
"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": 0.9985845163096919,
"total_bits": 532030464,
"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": 0.9991163368364429,
"total_bits": 552898560,
"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": 0.9995908544530577,
"total_bits": 708191232,
"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.40.block_sparse_moe": [
{
"accuracy": 0.9633940236740991,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9645016036535564,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9680106449676187,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9688297513578283,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9821942839281339,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9836469432339072,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9847329357933057,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910463352050436,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9917786357875326,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910339405515084,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921530309917503,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954540804518681,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.996099440589253,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975797885011783,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976709176614684,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978555782428501,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993476096684677,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.41.self_attn": [
{
"accuracy": 0.9810105923465208,
"total_bits": 187058176,
"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": 0.9822489647194743,
"total_bits": 192563200,
"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": 0.9847654203737253,
"total_bits": 197738240,
"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": 0.9880459285586288,
"total_bits": 234305024,
"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": 0.9904481551707968,
"total_bits": 276964352,
"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": 0.9906515453108832,
"total_bits": 277163264,
"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": 0.9942188829820799,
"total_bits": 355869696,
"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": 0.9944475062166978,
"total_bits": 356068608,
"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": 0.9948301943582728,
"total_bits": 359029248,
"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": 0.9951081453749028,
"total_bits": 364154880,
"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": 0.9951806116824675,
"total_bits": 365243648,
"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": 0.9954968747976971,
"total_bits": 368204288,
"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": 0.9961636997881884,
"total_bits": 373442560,
"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": 0.9965065011607581,
"total_bits": 378966016,
"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": 0.9978817708753539,
"total_bits": 458761216,
"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": 0.9982112011516859,
"total_bits": 467046400,
"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": 0.9984664453169037,
"total_bits": 532030464,
"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": 0.9989864925890671,
"total_bits": 552898560,
"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": 0.9995410437174876,
"total_bits": 708191232,
"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.41.block_sparse_moe": [
{
"accuracy": 0.9633973768275035,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9644940098453509,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.967986381955837,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9688005815995367,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9821949882040683,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9836356412913454,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9847151171661129,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910504100844264,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9917809181895695,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910399568463234,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921467811065284,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954635725038028,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960994048803848,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975919976989788,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.997678315677484,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.997862521712178,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993553532725568,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.42.self_attn": [
{
"accuracy": 0.9814556154765581,
"total_bits": 187058176,
"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": 0.9824838846137649,
"total_bits": 192563200,
"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": 0.9845736686532435,
"total_bits": 197738240,
"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": 0.9876984294847047,
"total_bits": 234305024,
"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": 0.9902881697359446,
"total_bits": 276964352,
"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": 0.9906651903630087,
"total_bits": 277163264,
"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": 0.9940840548002406,
"total_bits": 355869696,
"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": 0.9944188289074717,
"total_bits": 356068608,
"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": 0.9948757157636512,
"total_bits": 359029248,
"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": 0.9951305762055869,
"total_bits": 364154880,
"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": 0.9951612009960962,
"total_bits": 365243648,
"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": 0.9954870330463899,
"total_bits": 368204288,
"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": 0.9960709101103834,
"total_bits": 373442560,
"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": 0.9964128613324934,
"total_bits": 378966016,
"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": 0.9978505148380798,
"total_bits": 458761216,
"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": 0.9982086860691197,
"total_bits": 467046400,
"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": 0.9984454605544574,
"total_bits": 532030464,
"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": 0.9989866627637591,
"total_bits": 552898560,
"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": 0.9995504517987444,
"total_bits": 708191232,
"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.42.block_sparse_moe": [
{
"accuracy": 0.9627734561891932,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9638890129955191,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9673781630239988,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9681980102568081,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9818592357302183,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833280594627324,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9844054927020088,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9908696162318321,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9916202442867583,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9908743032107228,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9919995350861236,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953782348525956,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960256049801645,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975444845193507,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976351807984573,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9978190143408842,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993434376527531,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.43.self_attn": [
{
"accuracy": 0.9867265292100216,
"total_bits": 187058176,
"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": 0.9872132339269707,
"total_bits": 192563200,
"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": 0.9896266559432996,
"total_bits": 197738240,
"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": 0.9914331972893131,
"total_bits": 234305024,
"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": 0.9930076732447273,
"total_bits": 276964352,
"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": 0.993302901275456,
"total_bits": 277163264,
"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": 0.9956651210760403,
"total_bits": 355869696,
"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": 0.9959879782001831,
"total_bits": 356068608,
"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": 0.9963254328130892,
"total_bits": 359029248,
"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": 0.9964863277824694,
"total_bits": 364154880,
"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": 0.9965579183649664,
"total_bits": 365243648,
"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": 0.9967802473417434,
"total_bits": 368204288,
"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": 0.9974180149269829,
"total_bits": 373442560,
"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": 0.9976295305379225,
"total_bits": 378966016,
"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": 0.9985423332790991,
"total_bits": 458761216,
"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": 0.9987873752374741,
"total_bits": 467046400,
"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": 0.9988665973222325,
"total_bits": 532030464,
"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": 0.9993303538036631,
"total_bits": 552898560,
"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": 0.9996579147761957,
"total_bits": 708191232,
"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.43.block_sparse_moe": [
{
"accuracy": 0.9618947095188656,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9630421614764553,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9665429432337221,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9673639797654591,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9814090868458152,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9829278297134136,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9839981644855518,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906435877476868,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.991413628118799,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906346904800126,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918033134094194,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.995256601930841,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9959276113536601,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974810288986191,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975720394010607,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977530329549489,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993273977812159,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.44.self_attn": [
{
"accuracy": 0.9819898111745715,
"total_bits": 187058176,
"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": 0.9833161974031674,
"total_bits": 192563200,
"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": 0.9852371365812264,
"total_bits": 197738240,
"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": 0.9876609933317492,
"total_bits": 234305024,
"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": 0.9904498746373543,
"total_bits": 276964352,
"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": 0.9905858840793371,
"total_bits": 277163264,
"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": 0.9939408226575899,
"total_bits": 355869696,
"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": 0.9941665824061554,
"total_bits": 356068608,
"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": 0.994746850778986,
"total_bits": 359029248,
"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": 0.9950001234384743,
"total_bits": 364154880,
"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": 0.9951789002179315,
"total_bits": 365243648,
"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": 0.9955562583881578,
"total_bits": 368204288,
"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": 0.996267596504798,
"total_bits": 373442560,
"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": 0.9966461814964485,
"total_bits": 378966016,
"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": 0.9979578928629819,
"total_bits": 458761216,
"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": 0.9982898054754754,
"total_bits": 467046400,
"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": 0.9984511586616894,
"total_bits": 532030464,
"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": 0.9990074247624235,
"total_bits": 552898560,
"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": 0.999520711974535,
"total_bits": 708191232,
"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.44.block_sparse_moe": [
{
"accuracy": 0.9621487131440326,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.963305560185721,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9667633701311915,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9675722795195485,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9815427331057819,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9830436169316894,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9840895113731293,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9907157152499023,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9914737374913928,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9907136645756269,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918613627454952,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952986566472406,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9959587499156202,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975035903840571,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975952700233871,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9977717855761415,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993364390209439,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.45.self_attn": [
{
"accuracy": 0.9820341770292113,
"total_bits": 187058176,
"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": 0.982650142604191,
"total_bits": 192563200,
"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": 0.9848117201932167,
"total_bits": 197738240,
"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": 0.9879383378613152,
"total_bits": 234305024,
"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": 0.9906338457214204,
"total_bits": 276964352,
"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": 0.9909072242628195,
"total_bits": 277163264,
"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": 0.9941883640069711,
"total_bits": 355869696,
"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": 0.9945362708286235,
"total_bits": 356068608,
"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": 0.9949820628144631,
"total_bits": 359029248,
"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": 0.9952562814946041,
"total_bits": 364154880,
"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": 0.995305740678879,
"total_bits": 365243648,
"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": 0.9956650069767707,
"total_bits": 368204288,
"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": 0.996195265606634,
"total_bits": 373442560,
"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": 0.9965453806676363,
"total_bits": 378966016,
"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": 0.9978668917472916,
"total_bits": 458761216,
"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": 0.9982348061350518,
"total_bits": 467046400,
"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": 0.9984600140504833,
"total_bits": 532030464,
"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": 0.9990235074319093,
"total_bits": 552898560,
"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": 0.9995514504507331,
"total_bits": 708191232,
"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.45.block_sparse_moe": [
{
"accuracy": 0.9609313419971027,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9621149072129476,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9656374100595713,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9664634624496102,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9809163559326216,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9824806039564704,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9835494073647025,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9903989339884567,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9911880496350166,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9903942355395932,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915893220979917,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951345055535632,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958213444823694,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974141018710246,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975092586950938,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976888470425221,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993101831978971,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.46.self_attn": [
{
"accuracy": 0.9828161103650928,
"total_bits": 187058176,
"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": 0.9836451205866117,
"total_bits": 192563200,
"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": 0.9864156305005676,
"total_bits": 197738240,
"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": 0.9893660628432898,
"total_bits": 234305024,
"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": 0.9912127429178279,
"total_bits": 276964352,
"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": 0.9913065714859649,
"total_bits": 277163264,
"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": 0.9946020583033954,
"total_bits": 355869696,
"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": 0.9949385226998282,
"total_bits": 356068608,
"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": 0.9953731380795178,
"total_bits": 359029248,
"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": 0.9955968695055497,
"total_bits": 364154880,
"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": 0.995537791107046,
"total_bits": 365243648,
"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": 0.9958253737882172,
"total_bits": 368204288,
"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": 0.9965367731515711,
"total_bits": 373442560,
"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": 0.9968204880378356,
"total_bits": 378966016,
"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": 0.998053433202011,
"total_bits": 458761216,
"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": 0.9983921084234393,
"total_bits": 467046400,
"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": 0.9985672237265757,
"total_bits": 532030464,
"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": 0.9991183710915624,
"total_bits": 552898560,
"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": 0.9995766843509811,
"total_bits": 708191232,
"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.46.block_sparse_moe": [
{
"accuracy": 0.9606473482164898,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9618506855086276,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9653502452726427,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9661724069891007,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9807644281046171,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9823322564265445,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833912569539327,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9903147232620731,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9911097197205221,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9903159462356645,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915140153055912,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950952112037492,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957855806293848,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973933533336478,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974883498844543,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976664994610474,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9993040133543362,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.47.self_attn": [
{
"accuracy": 0.9837541837049159,
"total_bits": 187058176,
"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": 0.9844146306675515,
"total_bits": 192563200,
"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": 0.9871853489526793,
"total_bits": 197738240,
"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": 0.9896220124424681,
"total_bits": 234305024,
"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": 0.9914951047516967,
"total_bits": 276964352,
"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": 0.9916125933737739,
"total_bits": 277163264,
"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": 0.9950612764437928,
"total_bits": 355869696,
"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": 0.9952719274565185,
"total_bits": 356068608,
"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": 0.9956540077140457,
"total_bits": 359029248,
"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": 0.9958080969334236,
"total_bits": 364154880,
"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": 0.9958390923413006,
"total_bits": 365243648,
"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": 0.9960466980762583,
"total_bits": 368204288,
"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": 0.9968060627212062,
"total_bits": 373442560,
"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": 0.9970358620569306,
"total_bits": 378966016,
"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": 0.9982008880347406,
"total_bits": 458761216,
"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": 0.9985128692290011,
"total_bits": 467046400,
"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": 0.9986978859554878,
"total_bits": 532030464,
"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": 0.9992311590651394,
"total_bits": 552898560,
"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": 0.9996134875997852,
"total_bits": 708191232,
"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.47.block_sparse_moe": [
{
"accuracy": 0.9586610393108506,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9599289959786754,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9635555804952195,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9644101754223046,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9798007374629378,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9814426192131481,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9825282661048206,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9898324771539161,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906649727521366,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9898406018159891,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9910940723003525,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948573158686295,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9955786562549245,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972688542911783,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973690517638859,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975509121401334,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992758828846394,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.48.self_attn": [
{
"accuracy": 0.9806860387815457,
"total_bits": 187058176,
"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": 0.98176694999596,
"total_bits": 192563200,
"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": 0.9841849057395992,
"total_bits": 197738240,
"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": 0.9880739808033564,
"total_bits": 234305024,
"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": 0.9900459568751486,
"total_bits": 276964352,
"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": 0.9901916372687801,
"total_bits": 277163264,
"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": 0.9938754830999594,
"total_bits": 355869696,
"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": 0.9941778276850911,
"total_bits": 356068608,
"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": 0.9947190010846642,
"total_bits": 359029248,
"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": 0.9949795453760185,
"total_bits": 364154880,
"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": 0.9949972554060974,
"total_bits": 365243648,
"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": 0.9953075757029614,
"total_bits": 368204288,
"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": 0.9959665427082464,
"total_bits": 373442560,
"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": 0.9963017173454558,
"total_bits": 378966016,
"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": 0.9977692235751372,
"total_bits": 458761216,
"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": 0.9981342678901887,
"total_bits": 467046400,
"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": 0.9984080691907653,
"total_bits": 532030464,
"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": 0.9989749235315494,
"total_bits": 552898560,
"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": 0.9995386904660049,
"total_bits": 708191232,
"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.48.block_sparse_moe": [
{
"accuracy": 0.9567159464288699,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9580745199008992,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.961748117286908,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9626222976336354,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9788560029119253,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.980575759797112,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.981684148262598,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9893354001013857,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9902002712662675,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9893616298458686,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9906768758144033,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946128305807513,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953681931120196,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971288382443354,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972446154631478,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974315056552816,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9992323153051793,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.49.self_attn": [
{
"accuracy": 0.9795048190770965,
"total_bits": 187058176,
"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": 0.9804737558588386,
"total_bits": 192563200,
"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": 0.983666231953784,
"total_bits": 197738240,
"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": 0.9868544010739577,
"total_bits": 234305024,
"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": 0.9892307113071805,
"total_bits": 276964352,
"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": 0.9895554128789196,
"total_bits": 277163264,
"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": 0.9935379949349322,
"total_bits": 355869696,
"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": 0.993829883733078,
"total_bits": 356068608,
"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": 0.9943629301300174,
"total_bits": 359029248,
"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": 0.9946563394348088,
"total_bits": 364154880,
"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": 0.9947052421660996,
"total_bits": 365243648,
"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": 0.9951587223055723,
"total_bits": 368204288,
"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": 0.9959701661167568,
"total_bits": 373442560,
"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": 0.9963079026508096,
"total_bits": 378966016,
"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": 0.9977373455491799,
"total_bits": 458761216,
"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": 0.9981465902253005,
"total_bits": 467046400,
"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": 0.99828558220332,
"total_bits": 532030464,
"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": 0.9989499202224819,
"total_bits": 552898560,
"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": 0.9995193961750422,
"total_bits": 708191232,
"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.49.block_sparse_moe": [
{
"accuracy": 0.953174184125505,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9545918851110496,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9584693750856739,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9594088105582878,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9772567840685186,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.97908663357559,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9802879132330418,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884967981396537,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894393986385119,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9885419055231308,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9899550737617048,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941932477277556,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.995008737792408,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969022771172029,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.997028885960677,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972359246488562,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991672036688358,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.50.self_attn": [
{
"accuracy": 0.9735130575534544,
"total_bits": 187058176,
"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": 0.9750330219731519,
"total_bits": 192563200,
"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": 0.9789924479432797,
"total_bits": 197738240,
"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": 0.9832806635746046,
"total_bits": 234305024,
"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": 0.9866727777315598,
"total_bits": 276964352,
"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": 0.9868540813852298,
"total_bits": 277163264,
"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": 0.9923138671897744,
"total_bits": 355869696,
"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": 0.9926502219107198,
"total_bits": 356068608,
"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": 0.993233110559614,
"total_bits": 359029248,
"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": 0.9935420007248851,
"total_bits": 364154880,
"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": 0.9934176196540264,
"total_bits": 365243648,
"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": 0.993907599905996,
"total_bits": 368204288,
"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": 0.9948399832692781,
"total_bits": 373442560,
"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": 0.9953514119904292,
"total_bits": 378966016,
"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": 0.997136469060359,
"total_bits": 458761216,
"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": 0.9976340369260135,
"total_bits": 467046400,
"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": 0.9979616828556908,
"total_bits": 532030464,
"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": 0.9987215058767775,
"total_bits": 552898560,
"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": 0.9994278497807069,
"total_bits": 708191232,
"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.50.block_sparse_moe": [
{
"accuracy": 0.9524677952653483,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9540932239278367,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9577198032485812,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9586446414652624,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.97725523523006,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9791491074781669,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9803738942075717,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884904623766871,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894373878395479,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.988579193350712,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9899669561878239,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9942006285204307,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950080952656112,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968611124449557,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970248976268953,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972417787464923,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991146882377123,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.51.self_attn": [
{
"accuracy": 0.9676065205743438,
"total_bits": 187058176,
"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": 0.9702441683529239,
"total_bits": 192563200,
"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": 0.9745421240008191,
"total_bits": 197738240,
"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": 0.9798238403899105,
"total_bits": 234305024,
"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": 0.9842059281899741,
"total_bits": 276964352,
"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": 0.9845577367630444,
"total_bits": 277163264,
"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": 0.991050325677191,
"total_bits": 355869696,
"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": 0.9912906576585221,
"total_bits": 356068608,
"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": 0.9918501083494017,
"total_bits": 359029248,
"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": 0.9922382896520981,
"total_bits": 364154880,
"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": 0.9920987111369246,
"total_bits": 365243648,
"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": 0.9923363697813138,
"total_bits": 368204288,
"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": 0.9936466923375663,
"total_bits": 373442560,
"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": 0.9941641877062226,
"total_bits": 378966016,
"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": 0.9964880223361481,
"total_bits": 458761216,
"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": 0.9970496921971637,
"total_bits": 467046400,
"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": 0.997635343589967,
"total_bits": 532030464,
"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": 0.9984148408831587,
"total_bits": 552898560,
"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": 0.9993371100266047,
"total_bits": 708191232,
"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.51.block_sparse_moe": [
{
"accuracy": 0.9533462407753656,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9548407846963719,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.958680860796257,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9596241194950907,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.977242974662467,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9791347008492601,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.980352671963996,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884548559784889,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9893940758989438,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9885523567025206,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9899693227765199,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.994198262752769,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950131475435275,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968721115361213,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970381053127838,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972534520416766,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9991608560943094,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.52.self_attn": [
{
"accuracy": 0.966577016601437,
"total_bits": 187058176,
"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": 0.9684451403782556,
"total_bits": 192563200,
"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": 0.9722407895483469,
"total_bits": 197738240,
"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": 0.9770843927797518,
"total_bits": 234305024,
"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": 0.9843974904598374,
"total_bits": 276964352,
"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": 0.984561054722259,
"total_bits": 277163264,
"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": 0.9910306557522792,
"total_bits": 355869696,
"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": 0.9912685980089009,
"total_bits": 356068608,
"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": 0.9918215039039129,
"total_bits": 359029248,
"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": 0.9923194713743502,
"total_bits": 364154880,
"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": 0.9923480395227671,
"total_bits": 365243648,
"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": 0.9926758685492372,
"total_bits": 368204288,
"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": 0.9935331011289045,
"total_bits": 373442560,
"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": 0.9941278258706198,
"total_bits": 378966016,
"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": 0.9963141281378308,
"total_bits": 458761216,
"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": 0.9969594784840745,
"total_bits": 467046400,
"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": 0.9976280886361277,
"total_bits": 532030464,
"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": 0.998355223913677,
"total_bits": 552898560,
"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": 0.99932564874823,
"total_bits": 708191232,
"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.52.block_sparse_moe": [
{
"accuracy": 0.9518943795640218,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9534526469283983,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9574413414260274,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9584702037666973,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9763507550386222,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9782941397769671,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9796874880496609,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9878722584414247,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.988865764762618,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880446180652239,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9895294064312781,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.993935884319638,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947884750282882,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.99670576945707,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969039306988156,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971621795802524,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990880346127884,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.53.self_attn": [
{
"accuracy": 0.9648540285661033,
"total_bits": 187058176,
"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": 0.9668726566198625,
"total_bits": 192563200,
"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": 0.9724381554283594,
"total_bits": 197738240,
"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": 0.9785762897839672,
"total_bits": 234305024,
"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": 0.9825979721683421,
"total_bits": 276964352,
"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": 0.9830492968229871,
"total_bits": 277163264,
"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": 0.9899789944634234,
"total_bits": 355869696,
"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": 0.9903729506768286,
"total_bits": 356068608,
"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": 0.9910490511132306,
"total_bits": 359029248,
"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": 0.9914215354641017,
"total_bits": 364154880,
"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": 0.9914681601191038,
"total_bits": 365243648,
"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": 0.9920058880050323,
"total_bits": 368204288,
"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": 0.9931850538000857,
"total_bits": 373442560,
"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": 0.9937258675866025,
"total_bits": 378966016,
"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": 0.99622202116849,
"total_bits": 458761216,
"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": 0.996846009284797,
"total_bits": 467046400,
"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": 0.997375624462668,
"total_bits": 532030464,
"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": 0.9982846820432889,
"total_bits": 552898560,
"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": 0.9992545620556675,
"total_bits": 708191232,
"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.53.block_sparse_moe": [
{
"accuracy": 0.9483496229115286,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9498997562025723,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9541609963696254,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9553364244730849,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9747739373951366,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9768217344229159,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9784307222519266,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871028045339412,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9881904734739739,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9872709677897786,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9888399901652807,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935296964464023,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9944384792991179,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964999381244477,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966904616510299,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969966317416707,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990224538211662,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.54.self_attn": [
{
"accuracy": 0.9645984113020332,
"total_bits": 187058176,
"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": 0.9669630745131719,
"total_bits": 192563200,
"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": 0.9721537024567002,
"total_bits": 197738240,
"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": 0.9779169684962222,
"total_bits": 234305024,
"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": 0.9827172406704018,
"total_bits": 276964352,
"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": 0.9826562299736237,
"total_bits": 277163264,
"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": 0.9900030732939118,
"total_bits": 355869696,
"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": 0.9902610270257451,
"total_bits": 356068608,
"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": 0.9910711018662703,
"total_bits": 359029248,
"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": 0.9915460201568509,
"total_bits": 364154880,
"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": 0.9912421896465515,
"total_bits": 365243648,
"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": 0.9918396031042855,
"total_bits": 368204288,
"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": 0.9930282229018447,
"total_bits": 373442560,
"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": 0.9935964221919054,
"total_bits": 378966016,
"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": 0.996140098432079,
"total_bits": 458761216,
"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": 0.9967014038273575,
"total_bits": 467046400,
"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": 0.9973971601514342,
"total_bits": 532030464,
"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": 0.998281484102144,
"total_bits": 552898560,
"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": 0.999251259146197,
"total_bits": 708191232,
"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.54.block_sparse_moe": [
{
"accuracy": 0.9501394124603585,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9517448215108169,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9556711877843267,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9567273464053869,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.97556031328675,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775673720103345,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9790128229283973,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9873320039508766,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9884481625631452,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9875903701605765,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9891247663303817,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936786773737127,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9945777765793824,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965404106830982,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967701798121101,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970533381963737,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9990357963037432,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.55.self_attn": [
{
"accuracy": 0.9686907288667402,
"total_bits": 187058176,
"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": 0.9693894488246817,
"total_bits": 192563200,
"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": 0.9751748687244559,
"total_bits": 197738240,
"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": 0.9798559866061336,
"total_bits": 234305024,
"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": 0.9848972155075324,
"total_bits": 276964352,
"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": 0.9852368845849445,
"total_bits": 277163264,
"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": 0.9915342477248296,
"total_bits": 355869696,
"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": 0.9917126182515762,
"total_bits": 356068608,
"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": 0.9920253809541464,
"total_bits": 359029248,
"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": 0.9924891643765333,
"total_bits": 364154880,
"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": 0.9926816241472567,
"total_bits": 365243648,
"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": 0.9932007386202091,
"total_bits": 368204288,
"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": 0.9939251097507382,
"total_bits": 373442560,
"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": 0.9943134232522234,
"total_bits": 378966016,
"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": 0.9965940997935832,
"total_bits": 458761216,
"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": 0.997169707858886,
"total_bits": 467046400,
"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": 0.9976689619392058,
"total_bits": 532030464,
"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": 0.9984276747441312,
"total_bits": 552898560,
"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": 0.9992930593434721,
"total_bits": 708191232,
"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.55.block_sparse_moe": [
{
"accuracy": 0.9586245451906794,
"total_bits": 5403947904,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9599531113513207,
"total_bits": 5605274496,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9629756873963695,
"total_bits": 6255026176,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9639537825592255,
"total_bits": 7019438080,
"w1": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"w2": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.980578082417579,
"total_bits": 7906094592,
"w1": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9821997796626467,
"total_bits": 8128905216,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9834501590383681,
"total_bits": 8739823104,
"w1": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9896383772015964,
"total_bits": 9989670400,
"w1": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9907205705530941,
"total_bits": 10136928256,
"w1": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9901630545133039,
"total_bits": 10281119232,
"w1": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9914032227702831,
"total_bits": 10503929856,
"w1": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949098256980314,
"total_bits": 12656143872,
"w1": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956376715563238,
"total_bits": 12878954496,
"w1": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970886883828299,
"total_bits": 14664166912,
"w1": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.997357639808845,
"total_bits": 15165640704,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9976402391851145,
"total_bits": 16524595200,
"w1": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9989872008721393,
"total_bits": 19410276352,
"w1": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w3": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"w2": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.norm.norm": null,
"lm_head.linear": null
},
"last_module_idx": 114
}