{
"measurement": {
"model.layers.0.self_attn": [
{
"accuracy": 0.9177240133285522,
"total_bits": 86947200,
"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": 0.9217327833175659,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9360260963439941,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9545134902000427,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9605870246887207,
"total_bits": 129080000,
"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": 0.9651926159858704,
"total_bits": 129196400,
"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": 0.9716001749038696,
"total_bits": 165124800,
"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": 0.9770992398262024,
"total_bits": 165446000,
"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": 0.980593204498291,
"total_bits": 167052000,
"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": 0.9811851978302002,
"total_bits": 168979200,
"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": 0.9825040102005005,
"total_bits": 170156400,
"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": 0.9833828210830688,
"total_bits": 171148000,
"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": 0.9847335815429688,
"total_bits": 179721600,
"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": 0.9856932759284973,
"total_bits": 182291200,
"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": 0.99090176820755,
"total_bits": 220011200,
"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": 0.9911166429519653,
"total_bits": 223251200,
"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": 0.9924720525741577,
"total_bits": 247044800,
"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": 0.9930375814437866,
"total_bits": 271379200,
"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": 0.9951440691947937,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.0.mlp": [
{
"accuracy": 0.9365872740745544,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9379965662956238,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9532368779182434,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9582359194755554,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9715542197227478,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9732175469398499,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9801041483879089,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9847777485847473,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9861656427383423,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9855791926383972,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9869512915611267,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923688769340515,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9923235774040222,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950206279754639,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956074953079224,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965313673019409,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969714879989624,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.1.self_attn": [
{
"accuracy": 0.9129329919815063,
"total_bits": 86947200,
"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": 0.9170666933059692,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9275577068328857,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9523346424102783,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9577993154525757,
"total_bits": 129080000,
"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": 0.9586471319198608,
"total_bits": 129196400,
"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": 0.9717002511024475,
"total_bits": 165124800,
"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": 0.9727781414985657,
"total_bits": 165446000,
"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": 0.9755765199661255,
"total_bits": 167052000,
"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": 0.9765754342079163,
"total_bits": 168979200,
"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": 0.9784319996833801,
"total_bits": 170156400,
"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": 0.9795116782188416,
"total_bits": 171148000,
"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": 0.9808445572853088,
"total_bits": 179721600,
"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": 0.9819432497024536,
"total_bits": 182291200,
"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": 0.9880783557891846,
"total_bits": 220011200,
"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": 0.9889917969703674,
"total_bits": 223251200,
"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": 0.9900076389312744,
"total_bits": 247044800,
"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": 0.9918657541275024,
"total_bits": 271379200,
"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": 0.9940392374992371,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.1.mlp": [
{
"accuracy": 0.8965533971786499,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.8988513350486755,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9194796085357666,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.925621747970581,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9502027630805969,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9537826776504517,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9626827836036682,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9742567539215088,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9768008589744568,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9746943712234497,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9777461290359497,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871538281440735,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9879993200302124,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925115704536438,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9929497838020325,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943098425865173,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9959841966629028,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.2.self_attn": [
{
"accuracy": 0.9176256656646729,
"total_bits": 86947200,
"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": 0.9214654564857483,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9327325224876404,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9556199908256531,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9605461955070496,
"total_bits": 129080000,
"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": 0.961024284362793,
"total_bits": 129196400,
"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": 0.9764515161514282,
"total_bits": 165124800,
"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": 0.9770482778549194,
"total_bits": 165446000,
"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": 0.9791052341461182,
"total_bits": 167052000,
"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": 0.9799665212631226,
"total_bits": 168979200,
"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": 0.9802554249763489,
"total_bits": 170156400,
"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": 0.9815042018890381,
"total_bits": 171148000,
"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": 0.9834611415863037,
"total_bits": 179721600,
"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": 0.9846301674842834,
"total_bits": 182291200,
"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": 0.9907756447792053,
"total_bits": 220011200,
"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": 0.991156280040741,
"total_bits": 223251200,
"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": 0.9931634068489075,
"total_bits": 247044800,
"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": 0.9935560822486877,
"total_bits": 271379200,
"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": 0.9958291053771973,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.2.mlp": [
{
"accuracy": 0.904305100440979,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9106094241142273,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9213430881500244,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9244486689567566,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9601025581359863,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9621535539627075,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9674994945526123,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9778563380241394,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9797559976577759,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9804362654685974,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9824299812316895,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.98982173204422,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9902949929237366,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935250878334045,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9939770698547363,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947821497917175,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958178400993347,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.3.self_attn": [
{
"accuracy": 0.9432005286216736,
"total_bits": 86947200,
"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": 0.9445347785949707,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9540787935256958,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.964008629322052,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9678102135658264,
"total_bits": 129080000,
"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": 0.9698260426521301,
"total_bits": 129196400,
"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": 0.9740789532661438,
"total_bits": 165124800,
"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": 0.9765285849571228,
"total_bits": 165446000,
"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": 0.9798696637153625,
"total_bits": 167052000,
"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": 0.9803239703178406,
"total_bits": 168979200,
"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": 0.9849774837493896,
"total_bits": 170156400,
"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": 0.9865783452987671,
"total_bits": 171148000,
"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": 0.9864001870155334,
"total_bits": 179721600,
"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": 0.9880905151367188,
"total_bits": 182291200,
"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": 0.9922593832015991,
"total_bits": 220011200,
"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": 0.9935751557350159,
"total_bits": 223251200,
"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": 0.9932655692100525,
"total_bits": 247044800,
"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": 0.9964281916618347,
"total_bits": 271379200,
"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": 0.9974014163017273,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.3.mlp": [
{
"accuracy": 0.910824179649353,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9138386249542236,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9278630018234253,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9319697022438049,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9567090272903442,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9597564339637756,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9659252166748047,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.977672815322876,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9797810912132263,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9779419302940369,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9806693196296692,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9888927936553955,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.989818811416626,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936943650245667,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940344095230103,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950266480445862,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969090223312378,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.4.self_attn": [
{
"accuracy": 0.9314757585525513,
"total_bits": 86947200,
"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": 0.9325720071792603,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9502220153808594,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9595361351966858,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9654639959335327,
"total_bits": 129080000,
"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": 0.96713787317276,
"total_bits": 129196400,
"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": 0.9722180962562561,
"total_bits": 165124800,
"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": 0.9741367101669312,
"total_bits": 165446000,
"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": 0.977820098400116,
"total_bits": 167052000,
"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": 0.9783415794372559,
"total_bits": 168979200,
"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": 0.9828382134437561,
"total_bits": 170156400,
"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": 0.9847133159637451,
"total_bits": 171148000,
"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": 0.984296977519989,
"total_bits": 179721600,
"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": 0.9862185716629028,
"total_bits": 182291200,
"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": 0.9912256002426147,
"total_bits": 220011200,
"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": 0.9926730990409851,
"total_bits": 223251200,
"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": 0.9922121167182922,
"total_bits": 247044800,
"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": 0.995746374130249,
"total_bits": 271379200,
"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": 0.9964984655380249,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.4.mlp": [
{
"accuracy": 0.8992438316345215,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9017688035964966,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9177355766296387,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9224463701248169,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9497400522232056,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9534088373184204,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9602686166763306,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9740990996360779,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9765526652336121,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9744430184364319,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9776059985160828,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9871250987052917,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9883031845092773,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927500486373901,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9931350350379944,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9942410588264465,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9965601563453674,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.5.self_attn": [
{
"accuracy": 0.9312288761138916,
"total_bits": 86947200,
"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": 0.9332429766654968,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9466700553894043,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9577977657318115,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9638015627861023,
"total_bits": 129080000,
"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": 0.9652858376502991,
"total_bits": 129196400,
"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": 0.9716836810112,
"total_bits": 165124800,
"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": 0.9738436937332153,
"total_bits": 165446000,
"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": 0.9774815440177917,
"total_bits": 167052000,
"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": 0.97789067029953,
"total_bits": 168979200,
"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": 0.9820955991744995,
"total_bits": 170156400,
"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": 0.9838878512382507,
"total_bits": 171148000,
"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": 0.9839197993278503,
"total_bits": 179721600,
"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": 0.9857984781265259,
"total_bits": 182291200,
"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": 0.9909078478813171,
"total_bits": 220011200,
"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": 0.9923155307769775,
"total_bits": 223251200,
"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": 0.9921332001686096,
"total_bits": 247044800,
"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": 0.99552321434021,
"total_bits": 271379200,
"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": 0.9966526627540588,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.5.mlp": [
{
"accuracy": 0.9235719442367554,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9276253581047058,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9318480491638184,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9333574175834656,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9689197540283203,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9733375310897827,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9763890504837036,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9778809547424316,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9822179675102234,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9878314137458801,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9887891411781311,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934679865837097,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9935567378997803,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956947565078735,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961302876472473,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966464042663574,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9973414540290833,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.6.self_attn": [
{
"accuracy": 0.9703402519226074,
"total_bits": 86947200,
"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": 0.9710289239883423,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9760013222694397,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9803446531295776,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9847868084907532,
"total_bits": 129080000,
"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": 0.9852306246757507,
"total_bits": 129196400,
"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": 0.9894976019859314,
"total_bits": 165124800,
"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": 0.9901916980743408,
"total_bits": 165446000,
"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": 0.9913412928581238,
"total_bits": 167052000,
"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": 0.9916229248046875,
"total_bits": 168979200,
"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": 0.9925192594528198,
"total_bits": 170156400,
"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": 0.9932081699371338,
"total_bits": 171148000,
"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": 0.9936240315437317,
"total_bits": 179721600,
"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": 0.9942994713783264,
"total_bits": 182291200,
"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": 0.9964954853057861,
"total_bits": 220011200,
"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": 0.9969252347946167,
"total_bits": 223251200,
"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": 0.9972065091133118,
"total_bits": 247044800,
"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": 0.9980089664459229,
"total_bits": 271379200,
"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": 0.9987316131591797,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.6.mlp": [
{
"accuracy": 0.9520189166069031,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.953398585319519,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9604594111442566,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9625735878944397,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9761260151863098,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9778745770454407,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9809656143188477,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9875475168228149,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9886812567710876,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9876111149787903,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9889976978302002,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936081767082214,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940297603607178,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960625767707825,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9962632656097412,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966981410980225,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9975510835647583,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.7.self_attn": [
{
"accuracy": 0.9654545187950134,
"total_bits": 86947200,
"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": 0.9665160179138184,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9731841683387756,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9787471294403076,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9829807877540588,
"total_bits": 129080000,
"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": 0.9831844568252563,
"total_bits": 129196400,
"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": 0.990562379360199,
"total_bits": 165124800,
"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": 0.991041898727417,
"total_bits": 165446000,
"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": 0.9917930960655212,
"total_bits": 167052000,
"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": 0.9921669960021973,
"total_bits": 168979200,
"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": 0.9916236400604248,
"total_bits": 170156400,
"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": 0.9922457337379456,
"total_bits": 171148000,
"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": 0.9935548901557922,
"total_bits": 179721600,
"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": 0.9940813779830933,
"total_bits": 182291200,
"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": 0.9964919686317444,
"total_bits": 220011200,
"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": 0.996796727180481,
"total_bits": 223251200,
"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": 0.9975430965423584,
"total_bits": 247044800,
"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": 0.9978916645050049,
"total_bits": 271379200,
"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": 0.9987915754318237,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.7.mlp": [
{
"accuracy": 0.9469283223152161,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.948310136795044,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9564090967178345,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9588730335235596,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9734205007553101,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9753491878509521,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9788756966590881,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9862276911735535,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9875152111053467,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9864411950111389,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9881131649017334,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9931445121765137,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937621355056763,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960827827453613,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963220953941345,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9968886971473694,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9980957508087158,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.8.self_attn": [
{
"accuracy": 0.9616004228591919,
"total_bits": 86947200,
"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": 0.9633464813232422,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9695567488670349,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9759595990180969,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9811438322067261,
"total_bits": 129080000,
"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": 0.9813098907470703,
"total_bits": 129196400,
"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": 0.9898256063461304,
"total_bits": 165124800,
"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": 0.9901334643363953,
"total_bits": 165446000,
"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": 0.9909374117851257,
"total_bits": 167052000,
"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": 0.9913524389266968,
"total_bits": 168979200,
"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": 0.9906476140022278,
"total_bits": 170156400,
"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": 0.9913772344589233,
"total_bits": 171148000,
"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": 0.992758572101593,
"total_bits": 179721600,
"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": 0.9933127164840698,
"total_bits": 182291200,
"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": 0.9960657358169556,
"total_bits": 220011200,
"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": 0.996425449848175,
"total_bits": 223251200,
"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": 0.9973448514938354,
"total_bits": 247044800,
"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": 0.9977043867111206,
"total_bits": 271379200,
"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": 0.9987177848815918,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.8.mlp": [
{
"accuracy": 0.9434968829154968,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9449386596679688,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9536237716674805,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9562851786613464,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9716530442237854,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9737120270729065,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775218367576599,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9853124022483826,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9866890907287598,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9855623245239258,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9873430728912354,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9927249550819397,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933896064758301,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958850145339966,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961060881614685,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9967235922813416,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.998026430606842,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.9.self_attn": [
{
"accuracy": 0.9553995132446289,
"total_bits": 86947200,
"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": 0.9571937322616577,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9640243649482727,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9726303219795227,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9779067039489746,
"total_bits": 129080000,
"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": 0.9781882762908936,
"total_bits": 129196400,
"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": 0.9881770014762878,
"total_bits": 165124800,
"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": 0.9884291291236877,
"total_bits": 165446000,
"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": 0.9893790483474731,
"total_bits": 167052000,
"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": 0.9899032711982727,
"total_bits": 168979200,
"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": 0.9891637563705444,
"total_bits": 170156400,
"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": 0.989920973777771,
"total_bits": 171148000,
"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": 0.9914891719818115,
"total_bits": 179721600,
"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": 0.992179811000824,
"total_bits": 182291200,
"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": 0.9953780174255371,
"total_bits": 220011200,
"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": 0.9957698583602905,
"total_bits": 223251200,
"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": 0.996836245059967,
"total_bits": 247044800,
"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": 0.9972187876701355,
"total_bits": 271379200,
"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": 0.9983914494514465,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.9.mlp": [
{
"accuracy": 0.9422038197517395,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9436619877815247,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.952694833278656,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9555050134658813,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.970986008644104,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9731055498123169,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9770951271057129,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9849473834037781,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9863730669021606,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9852272868156433,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9870408177375793,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.992548942565918,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.993217945098877,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.99577796459198,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9960180521011353,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9966662526130676,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9979795813560486,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.10.self_attn": [
{
"accuracy": 0.9493499398231506,
"total_bits": 86947200,
"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": 0.9503210186958313,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9570291638374329,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9641897678375244,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9751027226448059,
"total_bits": 129080000,
"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": 0.9755076766014099,
"total_bits": 129196400,
"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": 0.9866898059844971,
"total_bits": 165124800,
"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": 0.9869493842124939,
"total_bits": 165446000,
"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": 0.988017737865448,
"total_bits": 167052000,
"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": 0.9887108206748962,
"total_bits": 168979200,
"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": 0.9876888394355774,
"total_bits": 170156400,
"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": 0.9885847568511963,
"total_bits": 171148000,
"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": 0.9899047613143921,
"total_bits": 179721600,
"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": 0.9906918406486511,
"total_bits": 182291200,
"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": 0.994533896446228,
"total_bits": 220011200,
"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": 0.9949538111686707,
"total_bits": 223251200,
"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": 0.9964499473571777,
"total_bits": 247044800,
"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": 0.9967092871665955,
"total_bits": 271379200,
"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": 0.9982918500900269,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.10.mlp": [
{
"accuracy": 0.9367719888687134,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9383903741836548,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9478542804718018,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9508419036865234,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.968203604221344,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9705210328102112,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9747180938720703,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9834887385368347,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9850257039070129,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9837997555732727,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.985788881778717,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9918237328529358,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9925611019134521,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9953529238700867,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9956453442573547,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9963309168815613,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.997803807258606,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.11.self_attn": [
{
"accuracy": 0.9467344880104065,
"total_bits": 86947200,
"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": 0.9483067393302917,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.957198441028595,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9659615159034729,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9739027619361877,
"total_bits": 129080000,
"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": 0.9739984273910522,
"total_bits": 129196400,
"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": 0.9854788184165955,
"total_bits": 165124800,
"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": 0.985880434513092,
"total_bits": 165446000,
"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": 0.9870201349258423,
"total_bits": 167052000,
"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": 0.9877007007598877,
"total_bits": 168979200,
"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": 0.9869197607040405,
"total_bits": 170156400,
"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": 0.9880322217941284,
"total_bits": 171148000,
"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": 0.9897122383117676,
"total_bits": 179721600,
"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": 0.9905611872673035,
"total_bits": 182291200,
"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": 0.9944283962249756,
"total_bits": 220011200,
"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": 0.9948906302452087,
"total_bits": 223251200,
"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": 0.9962124228477478,
"total_bits": 247044800,
"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": 0.9966403245925903,
"total_bits": 271379200,
"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": 0.9982412457466125,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.11.mlp": [
{
"accuracy": 0.9329814910888672,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.934683084487915,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.944854199886322,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9480962157249451,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9662190079689026,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9686933159828186,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9732176065444946,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9824193716049194,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9840959906578064,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9827898144721985,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9849174618721008,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.991330623626709,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9921249747276306,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950955510139465,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.995372474193573,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9961143732070923,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.997665524482727,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.12.self_attn": [
{
"accuracy": 0.9424950480461121,
"total_bits": 86947200,
"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": 0.9441617727279663,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9532767534255981,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9631575345993042,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9715968370437622,
"total_bits": 129080000,
"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": 0.9720828533172607,
"total_bits": 129196400,
"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": 0.9845325946807861,
"total_bits": 165124800,
"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": 0.9849380850791931,
"total_bits": 165446000,
"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": 0.9861252307891846,
"total_bits": 167052000,
"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": 0.9868570566177368,
"total_bits": 168979200,
"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": 0.985941469669342,
"total_bits": 170156400,
"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": 0.9868372082710266,
"total_bits": 171148000,
"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": 0.9886179566383362,
"total_bits": 179721600,
"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": 0.9895172119140625,
"total_bits": 182291200,
"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": 0.9938750267028809,
"total_bits": 220011200,
"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": 0.9943591356277466,
"total_bits": 223251200,
"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": 0.9959490895271301,
"total_bits": 247044800,
"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": 0.9963870048522949,
"total_bits": 271379200,
"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": 0.998094916343689,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.12.mlp": [
{
"accuracy": 0.928490161895752,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.930306077003479,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9413250088691711,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9448633193969727,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9639359712600708,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9665801525115967,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9715113639831543,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.981207013130188,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9829950332641602,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9816311001777649,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9838912487030029,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9907392263412476,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9915673136711121,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947534203529358,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950428009033203,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9958513975143433,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9974758625030518,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.13.self_attn": [
{
"accuracy": 0.939139723777771,
"total_bits": 86947200,
"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": 0.9412834048271179,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9504225850105286,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.961006760597229,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9697221517562866,
"total_bits": 129080000,
"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": 0.9703242778778076,
"total_bits": 129196400,
"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": 0.9834838509559631,
"total_bits": 165124800,
"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": 0.9840066432952881,
"total_bits": 165446000,
"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": 0.985402524471283,
"total_bits": 167052000,
"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": 0.9861466884613037,
"total_bits": 168979200,
"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": 0.9849755764007568,
"total_bits": 170156400,
"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": 0.9860799908638,
"total_bits": 171148000,
"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": 0.9879854917526245,
"total_bits": 179721600,
"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": 0.9889242649078369,
"total_bits": 182291200,
"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": 0.9935319423675537,
"total_bits": 220011200,
"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": 0.9940752387046814,
"total_bits": 223251200,
"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": 0.9956828355789185,
"total_bits": 247044800,
"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": 0.9961600303649902,
"total_bits": 271379200,
"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": 0.9979859590530396,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.13.mlp": [
{
"accuracy": 0.9238684177398682,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.925829291343689,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9373544454574585,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9411356449127197,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.961612343788147,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9643986821174622,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9696171879768372,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9799341559410095,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9818359613418579,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9804301261901855,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9828150272369385,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9900929927825928,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9909546971321106,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943125247955322,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9946519732475281,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9954921007156372,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971616268157959,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.14.self_attn": [
{
"accuracy": 0.934455156326294,
"total_bits": 86947200,
"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": 0.9367115497589111,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9466638565063477,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9578491449356079,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9673210978507996,
"total_bits": 129080000,
"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": 0.9675611257553101,
"total_bits": 129196400,
"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": 0.9822997450828552,
"total_bits": 165124800,
"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": 0.9826387166976929,
"total_bits": 165446000,
"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": 0.9842594861984253,
"total_bits": 167052000,
"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": 0.9850491285324097,
"total_bits": 168979200,
"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": 0.9836490750312805,
"total_bits": 170156400,
"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": 0.984866738319397,
"total_bits": 171148000,
"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": 0.9869013428688049,
"total_bits": 179721600,
"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": 0.98798006772995,
"total_bits": 182291200,
"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": 0.9929359555244446,
"total_bits": 220011200,
"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": 0.993538498878479,
"total_bits": 223251200,
"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": 0.9953737854957581,
"total_bits": 247044800,
"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": 0.9958711266517639,
"total_bits": 271379200,
"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": 0.9977965354919434,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.14.mlp": [
{
"accuracy": 0.9177756905555725,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.919827938079834,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9325910210609436,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9367133378982544,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9585274457931519,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9615681171417236,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9673069715499878,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9783830046653748,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9804465770721436,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9788675308227539,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9814786314964294,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9893354177474976,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9902874827384949,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9939250946044922,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9943070411682129,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.995246410369873,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970898032188416,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.15.self_attn": [
{
"accuracy": 0.9359668493270874,
"total_bits": 86947200,
"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": 0.9374649524688721,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9472782015800476,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.958419680595398,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9680261015892029,
"total_bits": 129080000,
"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": 0.9685199856758118,
"total_bits": 129196400,
"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": 0.9826374053955078,
"total_bits": 165124800,
"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": 0.983090877532959,
"total_bits": 165446000,
"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": 0.9844105839729309,
"total_bits": 167052000,
"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": 0.9852817058563232,
"total_bits": 168979200,
"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": 0.9840314984321594,
"total_bits": 170156400,
"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": 0.9851486682891846,
"total_bits": 171148000,
"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": 0.9870496988296509,
"total_bits": 179721600,
"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": 0.9880648255348206,
"total_bits": 182291200,
"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": 0.993003785610199,
"total_bits": 220011200,
"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": 0.9936627149581909,
"total_bits": 223251200,
"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": 0.9953904747962952,
"total_bits": 247044800,
"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": 0.9959225654602051,
"total_bits": 271379200,
"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": 0.997851550579071,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.15.mlp": [
{
"accuracy": 0.9138731956481934,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9160382151603699,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9294994473457336,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9338262677192688,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9565568566322327,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9597150087356567,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9657794833183289,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9773882031440735,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9795449376106262,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9778644442558289,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9805852174758911,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9888207316398621,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9898406267166138,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.993646502494812,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940595030784607,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950528740882874,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970109462738037,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.16.self_attn": [
{
"accuracy": 0.936841607093811,
"total_bits": 86947200,
"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": 0.9385818243026733,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9482191801071167,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9590591788291931,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9683783650398254,
"total_bits": 129080000,
"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": 0.9687911868095398,
"total_bits": 129196400,
"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": 0.9830402135848999,
"total_bits": 165124800,
"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": 0.9835072159767151,
"total_bits": 165446000,
"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": 0.9847487211227417,
"total_bits": 167052000,
"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": 0.9855918884277344,
"total_bits": 168979200,
"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": 0.9842639565467834,
"total_bits": 170156400,
"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": 0.9854537844657898,
"total_bits": 171148000,
"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": 0.9874510765075684,
"total_bits": 179721600,
"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": 0.9884675145149231,
"total_bits": 182291200,
"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": 0.9932616949081421,
"total_bits": 220011200,
"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": 0.9939048290252686,
"total_bits": 223251200,
"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": 0.9955838322639465,
"total_bits": 247044800,
"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": 0.9961603283882141,
"total_bits": 271379200,
"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": 0.9979615807533264,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.16.mlp": [
{
"accuracy": 0.9097747206687927,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9119685888290405,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9261963963508606,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9307481050491333,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9544898271560669,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9578119516372681,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9642164707183838,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9763457179069519,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9786041378974915,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9768284559249878,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9796774983406067,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9883002042770386,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9893888235092163,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933673739433289,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9937951564788818,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9948485493659973,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969077110290527,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.17.self_attn": [
{
"accuracy": 0.936918318271637,
"total_bits": 86947200,
"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": 0.9390618205070496,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9487131834030151,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9596605896949768,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9684903621673584,
"total_bits": 129080000,
"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": 0.9689411520957947,
"total_bits": 129196400,
"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": 0.9830051064491272,
"total_bits": 165124800,
"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": 0.9835382699966431,
"total_bits": 165446000,
"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": 0.985001266002655,
"total_bits": 167052000,
"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": 0.9857423901557922,
"total_bits": 168979200,
"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": 0.9843840599060059,
"total_bits": 170156400,
"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": 0.9856324791908264,
"total_bits": 171148000,
"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": 0.9875731468200684,
"total_bits": 179721600,
"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": 0.9886074662208557,
"total_bits": 182291200,
"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": 0.9933009743690491,
"total_bits": 220011200,
"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": 0.9939743280410767,
"total_bits": 223251200,
"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": 0.9956124424934387,
"total_bits": 247044800,
"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": 0.9962557554244995,
"total_bits": 271379200,
"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": 0.9980000257492065,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.17.mlp": [
{
"accuracy": 0.907934308052063,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9100937843322754,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9248316884040833,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9294832944869995,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9535717964172363,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9569474458694458,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9635227918624878,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9758857488632202,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9782149195671082,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9763720631599426,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9792813062667847,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880854487419128,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9892072081565857,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9932934641838074,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936941862106323,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.994781494140625,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969048500061035,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.18.self_attn": [
{
"accuracy": 0.9417886137962341,
"total_bits": 86947200,
"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": 0.9440310001373291,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9531508088111877,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.963138222694397,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9709886312484741,
"total_bits": 129080000,
"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": 0.9713147878646851,
"total_bits": 129196400,
"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": 0.9843199253082275,
"total_bits": 165124800,
"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": 0.9848012924194336,
"total_bits": 165446000,
"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": 0.9860446453094482,
"total_bits": 167052000,
"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": 0.9868277907371521,
"total_bits": 168979200,
"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": 0.9857326149940491,
"total_bits": 170156400,
"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": 0.9867305159568787,
"total_bits": 171148000,
"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": 0.9886807799339294,
"total_bits": 179721600,
"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": 0.9895712733268738,
"total_bits": 182291200,
"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": 0.993880033493042,
"total_bits": 220011200,
"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": 0.9944884181022644,
"total_bits": 223251200,
"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": 0.995948076248169,
"total_bits": 247044800,
"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": 0.9965692758560181,
"total_bits": 271379200,
"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": 0.9982197880744934,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.18.mlp": [
{
"accuracy": 0.9072564840316772,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9094761610031128,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9241164922714233,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9287437200546265,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9532315731048584,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9566372632980347,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9632073640823364,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9757593274116516,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9780802130699158,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9762300252914429,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.979153573513031,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9880441427230835,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9892010688781738,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9933152198791504,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9936983585357666,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947880506515503,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9969792366027832,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.19.self_attn": [
{
"accuracy": 0.9419925212860107,
"total_bits": 86947200,
"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": 0.9431938529014587,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9526045918464661,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9615346789360046,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9709963798522949,
"total_bits": 129080000,
"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": 0.9713408946990967,
"total_bits": 129196400,
"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": 0.9842761754989624,
"total_bits": 165124800,
"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": 0.9847032427787781,
"total_bits": 165446000,
"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": 0.986029326915741,
"total_bits": 167052000,
"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": 0.9867346286773682,
"total_bits": 168979200,
"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": 0.9856119751930237,
"total_bits": 170156400,
"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": 0.9867487549781799,
"total_bits": 171148000,
"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": 0.9887361526489258,
"total_bits": 179721600,
"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": 0.9896883368492126,
"total_bits": 182291200,
"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": 0.9938585758209229,
"total_bits": 220011200,
"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": 0.9944806694984436,
"total_bits": 223251200,
"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": 0.9958698153495789,
"total_bits": 247044800,
"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": 0.996433436870575,
"total_bits": 271379200,
"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": 0.997995913028717,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.19.mlp": [
{
"accuracy": 0.9091904759407043,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9113208651542664,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9255373477935791,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9300055503845215,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.954219400882721,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9575490951538086,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9639028906822205,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9762916564941406,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9785608053207397,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9767391085624695,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9796088337898254,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9883003830909729,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9894593954086304,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9934748411178589,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9938567280769348,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9949146509170532,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970992803573608,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.20.self_attn": [
{
"accuracy": 0.9486165642738342,
"total_bits": 86947200,
"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": 0.95090651512146,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9591884613037109,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.96675044298172,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9748371839523315,
"total_bits": 129080000,
"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": 0.9753091931343079,
"total_bits": 129196400,
"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": 0.9862366914749146,
"total_bits": 165124800,
"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": 0.9867591857910156,
"total_bits": 165446000,
"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": 0.987819492816925,
"total_bits": 167052000,
"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": 0.9883294701576233,
"total_bits": 168979200,
"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": 0.9875011444091797,
"total_bits": 170156400,
"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": 0.98850017786026,
"total_bits": 171148000,
"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": 0.9901918172836304,
"total_bits": 179721600,
"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": 0.9909708499908447,
"total_bits": 182291200,
"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": 0.9946736097335815,
"total_bits": 220011200,
"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": 0.9951024651527405,
"total_bits": 223251200,
"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": 0.9963586330413818,
"total_bits": 247044800,
"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": 0.9967754483222961,
"total_bits": 271379200,
"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": 0.9981701970100403,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.20.mlp": [
{
"accuracy": 0.9127780199050903,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9148420095443726,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9283568263053894,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9325607419013977,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.956044614315033,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9592171907424927,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9652310013771057,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9772341251373291,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9794236421585083,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9776419401168823,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9804047346115112,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9887553453445435,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9898816347122192,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.993736982345581,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9940869808197021,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950855374336243,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971990585327148,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.21.self_attn": [
{
"accuracy": 0.9539671540260315,
"total_bits": 86947200,
"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": 0.955508291721344,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9637141823768616,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.970506489276886,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9774361252784729,
"total_bits": 129080000,
"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": 0.9775776267051697,
"total_bits": 129196400,
"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": 0.9877943992614746,
"total_bits": 165124800,
"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": 0.9881541728973389,
"total_bits": 165446000,
"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": 0.989065408706665,
"total_bits": 167052000,
"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": 0.9896013140678406,
"total_bits": 168979200,
"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": 0.988885223865509,
"total_bits": 170156400,
"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": 0.9897370934486389,
"total_bits": 171148000,
"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": 0.9913874864578247,
"total_bits": 179721600,
"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": 0.9920946359634399,
"total_bits": 182291200,
"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": 0.9953165054321289,
"total_bits": 220011200,
"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": 0.9957361221313477,
"total_bits": 223251200,
"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": 0.9967994689941406,
"total_bits": 247044800,
"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": 0.9972267746925354,
"total_bits": 271379200,
"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": 0.998496949672699,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.21.mlp": [
{
"accuracy": 0.9138404130935669,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9159210324287415,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9288933873176575,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9329104423522949,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9565706849098206,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9597253203392029,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.965484619140625,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775335788726807,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.979685366153717,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9779269099235535,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9806569814682007,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9889184236526489,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9900363087654114,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9938618540763855,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941734075546265,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951335191726685,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9972708821296692,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.22.self_attn": [
{
"accuracy": 0.9498860836029053,
"total_bits": 86947200,
"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": 0.9516811370849609,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9599934220314026,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9675408601760864,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.975214421749115,
"total_bits": 129080000,
"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": 0.975623369216919,
"total_bits": 129196400,
"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": 0.9866546392440796,
"total_bits": 165124800,
"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": 0.987027645111084,
"total_bits": 165446000,
"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": 0.988040030002594,
"total_bits": 167052000,
"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": 0.9886026978492737,
"total_bits": 168979200,
"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": 0.9878209829330444,
"total_bits": 170156400,
"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": 0.9887345433235168,
"total_bits": 171148000,
"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": 0.9905011057853699,
"total_bits": 179721600,
"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": 0.9912907481193542,
"total_bits": 182291200,
"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": 0.9948492646217346,
"total_bits": 220011200,
"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": 0.9952542781829834,
"total_bits": 223251200,
"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": 0.9964652061462402,
"total_bits": 247044800,
"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": 0.996894359588623,
"total_bits": 271379200,
"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": 0.9983038902282715,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.22.mlp": [
{
"accuracy": 0.9151159524917603,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9171680808067322,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9296190738677979,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9334802031517029,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9571576118469238,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9603219032287598,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9658223390579224,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9777891039848328,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9799360632896423,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.978218138217926,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9809140563011169,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9890093803405762,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9901112914085388,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9938163161277771,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941934943199158,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9951046705245972,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9971760511398315,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.23.self_attn": [
{
"accuracy": 0.9577724933624268,
"total_bits": 86947200,
"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": 0.9591618180274963,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9664594531059265,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9726923108100891,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9794216156005859,
"total_bits": 129080000,
"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": 0.9796939492225647,
"total_bits": 129196400,
"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": 0.9888426065444946,
"total_bits": 165124800,
"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": 0.989107072353363,
"total_bits": 165446000,
"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": 0.9900698065757751,
"total_bits": 167052000,
"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": 0.9905381202697754,
"total_bits": 168979200,
"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": 0.9898517727851868,
"total_bits": 170156400,
"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": 0.9906265735626221,
"total_bits": 171148000,
"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": 0.992112398147583,
"total_bits": 179721600,
"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": 0.9927366971969604,
"total_bits": 182291200,
"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": 0.9957330226898193,
"total_bits": 220011200,
"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": 0.9960434436798096,
"total_bits": 223251200,
"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": 0.9970316886901855,
"total_bits": 247044800,
"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": 0.9973990321159363,
"total_bits": 271379200,
"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": 0.9985262751579285,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.23.mlp": [
{
"accuracy": 0.9153725504875183,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9174773097038269,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9294458627700806,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9331294298171997,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9573197364807129,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9605779051780701,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9658567309379578,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9778885245323181,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9800094366073608,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9783627986907959,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9810419678688049,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9890649914741516,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9900888204574585,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.993817925453186,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9941460490226746,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9950054883956909,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.997000515460968,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.24.self_attn": [
{
"accuracy": 0.9590924978256226,
"total_bits": 86947200,
"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": 0.9607964754104614,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9673714637756348,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9735159873962402,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9798654317855835,
"total_bits": 129080000,
"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": 0.9801721572875977,
"total_bits": 129196400,
"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": 0.9890934228897095,
"total_bits": 165124800,
"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": 0.9894006848335266,
"total_bits": 165446000,
"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": 0.9901682138442993,
"total_bits": 167052000,
"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": 0.9906661510467529,
"total_bits": 168979200,
"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": 0.9902133941650391,
"total_bits": 170156400,
"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": 0.9908740520477295,
"total_bits": 171148000,
"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": 0.9923003315925598,
"total_bits": 179721600,
"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": 0.9928999543190002,
"total_bits": 182291200,
"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": 0.9958410859107971,
"total_bits": 220011200,
"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": 0.9961578845977783,
"total_bits": 223251200,
"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": 0.997101366519928,
"total_bits": 247044800,
"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": 0.9974644184112549,
"total_bits": 271379200,
"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": 0.9986010789871216,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.24.mlp": [
{
"accuracy": 0.9115006327629089,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9144500494003296,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9261060357093811,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9296735525131226,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9563577175140381,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9590162634849548,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9641901850700378,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9768136143684387,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9784690737724304,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9775051474571228,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9802490472793579,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9885525703430176,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9893867373466492,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.993202269077301,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.993727445602417,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9945425987243652,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9964268803596497,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.25.self_attn": [
{
"accuracy": 0.9577867984771729,
"total_bits": 86947200,
"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": 0.9599529504776001,
"total_bits": 90224000,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9680209755897522,
"total_bits": 98709200,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9741918444633484,
"total_bits": 108832800,
"q_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"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": 0.9806426763534546,
"total_bits": 129080000,
"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": 0.9808921217918396,
"total_bits": 129196400,
"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": 0.9901913404464722,
"total_bits": 165124800,
"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": 0.9904928207397461,
"total_bits": 165446000,
"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": 0.991197943687439,
"total_bits": 167052000,
"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": 0.9915329217910767,
"total_bits": 168979200,
"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": 0.990847110748291,
"total_bits": 170156400,
"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": 0.9915829300880432,
"total_bits": 171148000,
"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": 0.9928789734840393,
"total_bits": 179721600,
"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": 0.993245005607605,
"total_bits": 182291200,
"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": 0.9960154891014099,
"total_bits": 220011200,
"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": 0.9962084293365479,
"total_bits": 223251200,
"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": 0.9971372485160828,
"total_bits": 247044800,
"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": 0.9972817897796631,
"total_bits": 271379200,
"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": 0.9982848167419434,
"total_bits": 328964800,
"q_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"k_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"v_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"o_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.layers.25.mlp": [
{
"accuracy": 0.934510350227356,
"total_bits": 185754976,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9364456534385681,
"total_bits": 192390496,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"6": 32,
"3": 64,
"2": 64
},
"bits": [
6,
3,
2
],
"bits_prop": [
0.05,
0.2,
0.75
],
"scale_bits": 4
}
},
{
"accuracy": 0.9440828561782837,
"total_bits": 214392160,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"3": 32
},
"bits": [
5,
3
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.94654381275177,
"total_bits": 240606560,
"gate_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"3": 64,
"2": 64
},
"bits": [
3,
2
],
"bits_prop": [
0.3,
0.7
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9676538705825806,
"total_bits": 272658944,
"gate_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 128,
"3": 128
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128,
"3": 128
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9699434041976929,
"total_bits": 278832160,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32,
"3": 32
},
"bits": [
8,
4,
3
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9734201431274414,
"total_bits": 299814944,
"gate_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32,
"3": 32
},
"bits": [
4,
3
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9829437732696533,
"total_bits": 342838544,
"gate_proj": {
"group_size": {
"4": 128
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 128
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9844988584518433,
"total_bits": 347900960,
"gate_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"4": 32
},
"bits": [
4
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"4": 32
},
"bits": [
8,
4
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9833990931510925,
"total_bits": 354169344,
"gate_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 128,
"4": 128
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 128,
"4": 128
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.985383152961731,
"total_bits": 360342560,
"gate_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"5": 32,
"4": 32
},
"bits": [
5,
4
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"5": 32,
"4": 32
},
"bits": [
8,
5,
4
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9913974404335022,
"total_bits": 435679744,
"gate_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128,
"5": 128
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128,
"5": 128
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9920594692230225,
"total_bits": 441852960,
"gate_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 32,
"5": 32
},
"bits": [
6,
5
],
"bits_prop": [
0.25,
0.75
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 32,
"5": 32
},
"bits": [
8,
6,
5
],
"bits_prop": [
0.05,
0.1,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9947201013565063,
"total_bits": 503263744,
"gate_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"6": 128
},
"bits": [
6
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 32,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.05,
0.95
],
"scale_bits": 4
}
},
{
"accuracy": 0.9952026009559631,
"total_bits": 522550304,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.15,
0.85
],
"scale_bits": 4
}
},
{
"accuracy": 0.9957330822944641,
"total_bits": 568835104,
"gate_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128,
"6": 128
},
"bits": [
8,
6
],
"bits_prop": [
0.1,
0.9
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
},
{
"accuracy": 0.9970856308937073,
"total_bits": 666156064,
"gate_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"up_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
},
"down_proj": {
"group_size": {
"8": 128
},
"bits": [
8
],
"bits_prop": [
1
],
"scale_bits": 4
}
}
],
"model.norm.norm": null,
"lm_head.linear": null
},
"last_module_idx": 54
}